gate apple gpu collectors + revert readme change

The previous behavior only caught some errors including inaccessible
hosts, but not others like failed authentication or service
unavailability. This largely applies when using a socket proxy and
having the retry mitigates some erroneous behavior.

.github/workflows/inactivity-actions.yml

@@ -6,6 +6,7 @@ on:
   workflow_dispatch:
 
 permissions:
+  actions: write
   issues: write
   pull-requests: write
 
@@ -48,6 +49,9 @@ jobs:
           # Action can not skip PRs, set it to 100 years to cover it.
           days-before-pr-stale: 36524
 
+          # Max issues to process before early exit. Next run resumes from cache. GH API limit: 5000.
+          operations-per-run: 1500
+
           # Labels
           stale-issue-label: 'stale'
           remove-stale-when-updated: true
@@ -56,4 +60,5 @@ jobs:
 
           # Exemptions
           exempt-assignees: true
-          exempt-milestones: true
+
+          exempt-milestones: true

.gitignore

@@ -10,6 +10,7 @@ dist
 *.exe
 internal/cmd/hub/hub
 internal/cmd/agent/agent
+agent.test
 node_modules
 build
 *timestamp*

Makefile

@@ -3,6 +3,40 @@ OS ?= $(shell go env GOOS)
 ARCH ?= $(shell go env GOARCH)
 # Skip building the web UI if true
 SKIP_WEB ?= false
+# Controls NVML/glibc agent build tag behavior:
+# - auto (default): enable on linux/amd64 glibc hosts
+# - true: always enable
+# - false: always disable
+NVML ?= auto
+
+# Detect glibc host for local linux/amd64 builds.
+HOST_GLIBC := $(shell \
+	if [ "$(OS)" = "linux" ] && [ "$(ARCH)" = "amd64" ]; then \
+		for p in /lib64/ld-linux-x86-64.so.2 /lib/x86_64-linux-gnu/ld-linux-x86-64.so.2 /lib/ld-linux-x86-64.so.2; do \
+			[ -e "$$p" ] && { echo true; exit 0; }; \
+		done; \
+		if command -v ldd >/dev/null 2>&1; then \
+			if ldd --version 2>&1 | tr '[:upper:]' '[:lower:]' | awk '/gnu libc|glibc/{found=1} END{exit !found}'; then \
+				echo true; \
+			else \
+				echo false; \
+			fi; \
+		else \
+			echo false; \
+		fi; \
+	else \
+		echo false; \
+	fi)
+
+# Enable glibc build tag for NVML on supported Linux builds.
+AGENT_GO_TAGS :=
+ifeq ($(NVML),true)
+AGENT_GO_TAGS := -tags glibc
+else ifeq ($(NVML),auto)
+ifeq ($(HOST_GLIBC),true)
+AGENT_GO_TAGS := -tags glibc
+endif
+endif
 
 # Set executable extension based on target OS
 EXE_EXT := $(if $(filter windows,$(OS)),.exe,)
@@ -54,7 +88,7 @@ fetch-smartctl-conditional:
 
 # Update build-agent to include conditional .NET build
 build-agent: tidy build-dotnet-conditional fetch-smartctl-conditional
-	GOOS=$(OS) GOARCH=$(ARCH) go build -o ./build/beszel-agent_$(OS)_$(ARCH)$(EXE_EXT) -ldflags "-w -s" ./internal/cmd/agent
+	GOOS=$(OS) GOARCH=$(ARCH) go build $(AGENT_GO_TAGS) -o ./build/beszel-agent_$(OS)_$(ARCH)$(EXE_EXT) -ldflags "-w -s" ./internal/cmd/agent
 
 build-hub: tidy $(if $(filter false,$(SKIP_WEB)),build-web-ui)
 	GOOS=$(OS) GOARCH=$(ARCH) go build -o ./build/beszel_$(OS)_$(ARCH)$(EXE_EXT) -ldflags "-w -s" ./internal/cmd/hub
@@ -90,9 +124,9 @@ dev-hub:
 
 dev-agent:
 	@if command -v entr >/dev/null 2>&1; then \
-		find ./internal/cmd/agent/*.go ./agent/*.go | entr -r go run github.com/henrygd/beszel/internal/cmd/agent; \
+		find ./internal/cmd/agent/*.go ./agent/*.go | entr -r go run $(AGENT_GO_TAGS) github.com/henrygd/beszel/internal/cmd/agent; \
 	else \
-		go run github.com/henrygd/beszel/internal/cmd/agent; \
+		go run $(AGENT_GO_TAGS) github.com/henrygd/beszel/internal/cmd/agent; \
 	fi
 	
 build-dotnet:

agent/docker.go

@@ -72,6 +72,7 @@ type dockerManager struct {
 	// cacheTimeMs -> DeltaTracker for network bytes sent/received
 	networkSentTrackers map[uint16]*deltatracker.DeltaTracker[string, uint64]
 	networkRecvTrackers map[uint16]*deltatracker.DeltaTracker[string, uint64]
+	retrySleep          func(time.Duration)
 }
 
 // userAgentRoundTripper is a custom http.RoundTripper that adds a User-Agent header to all requests
@@ -565,6 +566,7 @@ func newDockerManager() *dockerManager {
 		lastCpuReadTime:     make(map[uint16]map[string]time.Time),
 		networkSentTrackers: make(map[uint16]*deltatracker.DeltaTracker[string, uint64]),
 		networkRecvTrackers: make(map[uint16]*deltatracker.DeltaTracker[string, uint64]),
+		retrySleep:          time.Sleep,
 	}
 
 	// If using podman, return client
@@ -574,7 +576,7 @@ func newDockerManager() *dockerManager {
 		return manager
 	}
 
-	// this can take up to 5 seconds with retry, so run in goroutine
+	// run version check in goroutine to avoid blocking (server may not be ready and requires retries)
 	go manager.checkDockerVersion()
 
 	// give version check a chance to complete before returning
@@ -594,18 +596,18 @@ func (dm *dockerManager) checkDockerVersion() {
 	const versionMaxTries = 2
 	for i := 1; i <= versionMaxTries; i++ {
 		resp, err = dm.client.Get("http://localhost/version")
-		if err == nil {
+		if err == nil && resp.StatusCode == http.StatusOK {
 			break
 		}
 		if resp != nil {
 			resp.Body.Close()
 		}
 		if i < versionMaxTries {
-			slog.Debug("Failed to get Docker version; retrying", "attempt", i, "error", err)
-			time.Sleep(5 * time.Second)
+			slog.Debug("Failed to get Docker version; retrying", "attempt", i, "err", err, "response", resp)
+			dm.retrySleep(5 * time.Second)
 		}
 	}
-	if err != nil {
+	if err != nil || resp.StatusCode != http.StatusOK {
 		return
 	}
 	if err := dm.decode(resp, &versionInfo); err != nil {

agent/docker_test.go

@@ -5,7 +5,13 @@ package agent
 
 import (
 	"bytes"
+	"context"
 	"encoding/json"
+	"errors"
+	"fmt"
+	"net"
+	"net/http"
+	"net/http/httptest"
 	"os"
 	"strings"
 	"testing"
@@ -379,6 +385,117 @@ func TestDockerManagerCreation(t *testing.T) {
 	assert.NotNil(t, dm.networkRecvTrackers)
 }
 
+func TestCheckDockerVersion(t *testing.T) {
+	tests := []struct {
+		name      string
+		responses []struct {
+			statusCode int
+			body       string
+		}
+		expectedGood     bool
+		expectedRequests int
+	}{
+		{
+			name: "200 with good version on first try",
+			responses: []struct {
+				statusCode int
+				body       string
+			}{
+				{http.StatusOK, `{"Version":"25.0.1"}`},
+			},
+			expectedGood:     true,
+			expectedRequests: 1,
+		},
+		{
+			name: "200 with old version on first try",
+			responses: []struct {
+				statusCode int
+				body       string
+			}{
+				{http.StatusOK, `{"Version":"24.0.7"}`},
+			},
+			expectedGood:     false,
+			expectedRequests: 1,
+		},
+		{
+			name: "non-200 then 200 with good version",
+			responses: []struct {
+				statusCode int
+				body       string
+			}{
+				{http.StatusServiceUnavailable, `"not ready"`},
+				{http.StatusOK, `{"Version":"25.1.0"}`},
+			},
+			expectedGood:     true,
+			expectedRequests: 2,
+		},
+		{
+			name: "non-200 on all retries",
+			responses: []struct {
+				statusCode int
+				body       string
+			}{
+				{http.StatusInternalServerError, `"error"`},
+				{http.StatusUnauthorized, `"error"`},
+			},
+			expectedGood:     false,
+			expectedRequests: 2,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			requestCount := 0
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				idx := requestCount
+				requestCount++
+				if idx >= len(tt.responses) {
+					idx = len(tt.responses) - 1
+				}
+				w.WriteHeader(tt.responses[idx].statusCode)
+				fmt.Fprint(w, tt.responses[idx].body)
+			}))
+			defer server.Close()
+
+			dm := &dockerManager{
+				client: &http.Client{
+					Transport: &http.Transport{
+						DialContext: func(_ context.Context, network, _ string) (net.Conn, error) {
+							return net.Dial(network, server.Listener.Addr().String())
+						},
+					},
+				},
+				retrySleep: func(time.Duration) {},
+			}
+
+			dm.checkDockerVersion()
+
+			assert.Equal(t, tt.expectedGood, dm.goodDockerVersion)
+			assert.Equal(t, tt.expectedRequests, requestCount)
+		})
+	}
+
+	t.Run("request error on all retries", func(t *testing.T) {
+		requestCount := 0
+		dm := &dockerManager{
+			client: &http.Client{
+				Transport: &http.Transport{
+					DialContext: func(_ context.Context, _, _ string) (net.Conn, error) {
+						requestCount++
+						return nil, errors.New("connection refused")
+					},
+				},
+			},
+			retrySleep: func(time.Duration) {},
+		}
+
+		dm.checkDockerVersion()
+
+		assert.False(t, dm.goodDockerVersion)
+		assert.Equal(t, 2, requestCount)
+	})
+}
+
 func TestCycleCpuDeltas(t *testing.T) {
 	dm := &dockerManager{
 		lastCpuContainer: map[uint16]map[string]uint64{

agent/emmc_common.go

@@ -0,0 +1,95 @@
+package agent
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+)
+
+func isEmmcBlockName(name string) bool {
+	if !strings.HasPrefix(name, "mmcblk") {
+		return false
+	}
+	suffix := strings.TrimPrefix(name, "mmcblk")
+	if suffix == "" {
+		return false
+	}
+	for _, c := range suffix {
+		if c < '0' || c > '9' {
+			return false
+		}
+	}
+	return true
+}
+
+func parseHexOrDecByte(s string) (uint8, bool) {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return 0, false
+	}
+	base := 10
+	if strings.HasPrefix(s, "0x") || strings.HasPrefix(s, "0X") {
+		base = 16
+		s = s[2:]
+	}
+	parsed, err := strconv.ParseUint(s, base, 8)
+	if err != nil {
+		return 0, false
+	}
+	return uint8(parsed), true
+}
+
+func parseHexBytePair(s string) (uint8, uint8, bool) {
+	fields := strings.Fields(s)
+	if len(fields) < 2 {
+		return 0, 0, false
+	}
+	a, okA := parseHexOrDecByte(fields[0])
+	b, okB := parseHexOrDecByte(fields[1])
+	if !okA && !okB {
+		return 0, 0, false
+	}
+	return a, b, true
+}
+
+func emmcSmartStatus(preEOL uint8) string {
+	switch preEOL {
+	case 0x01:
+		return "PASSED"
+	case 0x02:
+		return "WARNING"
+	case 0x03:
+		return "FAILED"
+	default:
+		return "UNKNOWN"
+	}
+}
+
+func emmcPreEOLString(preEOL uint8) string {
+	switch preEOL {
+	case 0x01:
+		return "0x01 (normal)"
+	case 0x02:
+		return "0x02 (warning)"
+	case 0x03:
+		return "0x03 (urgent)"
+	default:
+		return fmt.Sprintf("0x%02x", preEOL)
+	}
+}
+
+func emmcLifeTimeString(v uint8) string {
+	// JEDEC eMMC: 0x01..0x0A => 0-100% used in 10% steps, 0x0B => exceeded.
+	switch {
+	case v == 0:
+		return "0x00 (not reported)"
+	case v >= 0x01 && v <= 0x0A:
+		low := int(v-1) * 10
+		high := int(v) * 10
+		return fmt.Sprintf("0x%02x (%d-%d%% used)", v, low, high)
+	case v == 0x0B:
+		return "0x0b (>100% used)"
+	default:
+		return fmt.Sprintf("0x%02x", v)
+	}
+}

agent/emmc_common_test.go

@@ -0,0 +1,78 @@
+package agent
+
+import "testing"
+
+func TestParseHexOrDecByte(t *testing.T) {
+	tests := []struct {
+		in   string
+		want uint8
+		ok   bool
+	}{
+		{"0x01", 1, true},
+		{"0X0b", 11, true},
+		{"01", 1, true},
+		{" 3 ", 3, true},
+		{"", 0, false},
+		{"0x", 0, false},
+		{"nope", 0, false},
+	}
+
+	for _, tt := range tests {
+		got, ok := parseHexOrDecByte(tt.in)
+		if ok != tt.ok || got != tt.want {
+			t.Fatalf("parseHexOrDecByte(%q) = (%d,%v), want (%d,%v)", tt.in, got, ok, tt.want, tt.ok)
+		}
+	}
+}
+
+func TestParseHexBytePair(t *testing.T) {
+	a, b, ok := parseHexBytePair("0x01 0x02\n")
+	if !ok || a != 1 || b != 2 {
+		t.Fatalf("parseHexBytePair hex = (%d,%d,%v), want (1,2,true)", a, b, ok)
+	}
+
+	a, b, ok = parseHexBytePair("01 02")
+	if !ok || a != 1 || b != 2 {
+		t.Fatalf("parseHexBytePair dec = (%d,%d,%v), want (1,2,true)", a, b, ok)
+	}
+
+	_, _, ok = parseHexBytePair("0x01")
+	if ok {
+		t.Fatalf("parseHexBytePair short input ok=true, want false")
+	}
+}
+
+func TestEmmcSmartStatus(t *testing.T) {
+	if got := emmcSmartStatus(0x01); got != "PASSED" {
+		t.Fatalf("emmcSmartStatus(0x01) = %q, want PASSED", got)
+	}
+	if got := emmcSmartStatus(0x02); got != "WARNING" {
+		t.Fatalf("emmcSmartStatus(0x02) = %q, want WARNING", got)
+	}
+	if got := emmcSmartStatus(0x03); got != "FAILED" {
+		t.Fatalf("emmcSmartStatus(0x03) = %q, want FAILED", got)
+	}
+	if got := emmcSmartStatus(0x00); got != "UNKNOWN" {
+		t.Fatalf("emmcSmartStatus(0x00) = %q, want UNKNOWN", got)
+	}
+}
+
+func TestIsEmmcBlockName(t *testing.T) {
+	cases := []struct {
+		name string
+		ok   bool
+	}{
+		{"mmcblk0", true},
+		{"mmcblk1", true},
+		{"mmcblk10", true},
+		{"mmcblk0p1", false},
+		{"sda", false},
+		{"mmcblk", false},
+		{"mmcblkA", false},
+	}
+	for _, c := range cases {
+		if got := isEmmcBlockName(c.name); got != c.ok {
+			t.Fatalf("isEmmcBlockName(%q) = %v, want %v", c.name, got, c.ok)
+		}
+	}
+}

agent/emmc_linux.go

@@ -0,0 +1,227 @@
+//go:build linux
+
+package agent
+
+import (
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"github.com/henrygd/beszel/internal/entities/smart"
+)
+
+// emmcSysfsRoot is a test hook; production value is "/sys".
+var emmcSysfsRoot = "/sys"
+
+type emmcHealth struct {
+	model    string
+	serial   string
+	revision string
+	capacity uint64
+	preEOL   uint8
+	lifeA    uint8
+	lifeB    uint8
+}
+
+func scanEmmcDevices() []*DeviceInfo {
+	blockDir := filepath.Join(emmcSysfsRoot, "class", "block")
+	entries, err := os.ReadDir(blockDir)
+	if err != nil {
+		return nil
+	}
+
+	devices := make([]*DeviceInfo, 0, 2)
+	for _, ent := range entries {
+		name := ent.Name()
+		if !isEmmcBlockName(name) {
+			continue
+		}
+
+		deviceDir := filepath.Join(blockDir, name, "device")
+		if !hasEmmcHealthFiles(deviceDir) {
+			continue
+		}
+
+		devPath := filepath.Join("/dev", name)
+		devices = append(devices, &DeviceInfo{
+			Name:     devPath,
+			Type:     "emmc",
+			InfoName: devPath + " [eMMC]",
+			Protocol: "MMC",
+		})
+	}
+
+	return devices
+}
+
+func (sm *SmartManager) collectEmmcHealth(deviceInfo *DeviceInfo) (bool, error) {
+	if deviceInfo == nil || deviceInfo.Name == "" {
+		return false, nil
+	}
+
+	base := filepath.Base(deviceInfo.Name)
+	if !isEmmcBlockName(base) && !strings.EqualFold(deviceInfo.Type, "emmc") && !strings.EqualFold(deviceInfo.Type, "mmc") {
+		return false, nil
+	}
+
+	health, ok := readEmmcHealth(base)
+	if !ok {
+		return false, nil
+	}
+
+	// Normalize the device type to keep pruning logic stable across refreshes.
+	deviceInfo.Type = "emmc"
+
+	key := health.serial
+	if key == "" {
+		key = filepath.Join("/dev", base)
+	}
+
+	status := emmcSmartStatus(health.preEOL)
+
+	attrs := []*smart.SmartAttribute{
+		{
+			Name:      "PreEOLInfo",
+			RawValue:  uint64(health.preEOL),
+			RawString: emmcPreEOLString(health.preEOL),
+		},
+		{
+			Name:      "DeviceLifeTimeEstA",
+			RawValue:  uint64(health.lifeA),
+			RawString: emmcLifeTimeString(health.lifeA),
+		},
+		{
+			Name:      "DeviceLifeTimeEstB",
+			RawValue:  uint64(health.lifeB),
+			RawString: emmcLifeTimeString(health.lifeB),
+		},
+	}
+
+	sm.Lock()
+	defer sm.Unlock()
+
+	if _, exists := sm.SmartDataMap[key]; !exists {
+		sm.SmartDataMap[key] = &smart.SmartData{}
+	}
+
+	data := sm.SmartDataMap[key]
+	data.ModelName = health.model
+	data.SerialNumber = health.serial
+	data.FirmwareVersion = health.revision
+	data.Capacity = health.capacity
+	data.Temperature = 0
+	data.SmartStatus = status
+	data.DiskName = filepath.Join("/dev", base)
+	data.DiskType = "emmc"
+	data.Attributes = attrs
+
+	return true, nil
+}
+
+func readEmmcHealth(blockName string) (emmcHealth, bool) {
+	var out emmcHealth
+
+	if !isEmmcBlockName(blockName) {
+		return out, false
+	}
+
+	deviceDir := filepath.Join(emmcSysfsRoot, "class", "block", blockName, "device")
+	preEOL, okPre := readHexByteFile(filepath.Join(deviceDir, "pre_eol_info"))
+
+	// Some kernels expose EXT_CSD lifetime via "life_time" (two bytes), others as
+	// separate files. Support both.
+	lifeA, lifeB, okLife := readLifeTime(deviceDir)
+
+	if !okPre && !okLife {
+		return out, false
+	}
+
+	out.preEOL = preEOL
+	out.lifeA = lifeA
+	out.lifeB = lifeB
+
+	out.model = readStringFile(filepath.Join(deviceDir, "name"))
+	out.serial = readStringFile(filepath.Join(deviceDir, "serial"))
+	out.revision = readStringFile(filepath.Join(deviceDir, "prv"))
+
+	if capBytes, ok := readBlockCapacityBytes(blockName); ok {
+		out.capacity = capBytes
+	}
+
+	return out, true
+}
+
+func readLifeTime(deviceDir string) (uint8, uint8, bool) {
+	if content, ok := readStringFileOK(filepath.Join(deviceDir, "life_time")); ok {
+		a, b, ok := parseHexBytePair(content)
+		return a, b, ok
+	}
+
+	a, okA := readHexByteFile(filepath.Join(deviceDir, "device_life_time_est_typ_a"))
+	b, okB := readHexByteFile(filepath.Join(deviceDir, "device_life_time_est_typ_b"))
+	if okA || okB {
+		return a, b, true
+	}
+	return 0, 0, false
+}
+
+func readBlockCapacityBytes(blockName string) (uint64, bool) {
+	sizePath := filepath.Join(emmcSysfsRoot, "class", "block", blockName, "size")
+	lbsPath := filepath.Join(emmcSysfsRoot, "class", "block", blockName, "queue", "logical_block_size")
+
+	sizeStr, ok := readStringFileOK(sizePath)
+	if !ok {
+		return 0, false
+	}
+	sectors, err := strconv.ParseUint(sizeStr, 10, 64)
+	if err != nil || sectors == 0 {
+		return 0, false
+	}
+
+	lbsStr, ok := readStringFileOK(lbsPath)
+	logicalBlockSize := uint64(512)
+	if ok {
+		if parsed, err := strconv.ParseUint(lbsStr, 10, 64); err == nil && parsed > 0 {
+			logicalBlockSize = parsed
+		}
+	}
+
+	return sectors * logicalBlockSize, true
+}
+
+func readHexByteFile(path string) (uint8, bool) {
+	content, ok := readStringFileOK(path)
+	if !ok {
+		return 0, false
+	}
+	b, ok := parseHexOrDecByte(content)
+	return b, ok
+}
+
+func readStringFile(path string) string {
+	content, _ := readStringFileOK(path)
+	return content
+}
+
+func readStringFileOK(path string) (string, bool) {
+	b, err := os.ReadFile(path)
+	if err != nil {
+		return "", false
+	}
+	return strings.TrimSpace(string(b)), true
+}
+
+func hasEmmcHealthFiles(deviceDir string) bool {
+	entries, err := os.ReadDir(deviceDir)
+	if err != nil {
+		return false
+	}
+	for _, ent := range entries {
+		switch ent.Name() {
+		case "pre_eol_info", "life_time", "device_life_time_est_typ_a", "device_life_time_est_typ_b":
+			return true
+		}
+	}
+	return false
+}

agent/emmc_linux_test.go

@@ -0,0 +1,80 @@
+//go:build linux
+
+package agent
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/smart"
+)
+
+func TestEmmcMockSysfsScanAndCollect(t *testing.T) {
+	tmp := t.TempDir()
+	prev := emmcSysfsRoot
+	emmcSysfsRoot = tmp
+	t.Cleanup(func() { emmcSysfsRoot = prev })
+
+	// Fake: /sys/class/block/mmcblk0
+	mmcDeviceDir := filepath.Join(tmp, "class", "block", "mmcblk0", "device")
+	mmcQueueDir := filepath.Join(tmp, "class", "block", "mmcblk0", "queue")
+	if err := os.MkdirAll(mmcDeviceDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.MkdirAll(mmcQueueDir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	write := func(path, content string) {
+		t.Helper()
+		if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	write(filepath.Join(mmcDeviceDir, "pre_eol_info"), "0x02\n")
+	write(filepath.Join(mmcDeviceDir, "life_time"), "0x04 0x05\n")
+	write(filepath.Join(mmcDeviceDir, "name"), "H26M52103FMR\n")
+	write(filepath.Join(mmcDeviceDir, "serial"), "01234567\n")
+	write(filepath.Join(mmcDeviceDir, "prv"), "0x08\n")
+	write(filepath.Join(mmcQueueDir, "logical_block_size"), "512\n")
+	write(filepath.Join(tmp, "class", "block", "mmcblk0", "size"), "1024\n") // sectors
+
+	devs := scanEmmcDevices()
+	if len(devs) != 1 {
+		t.Fatalf("scanEmmcDevices() = %d devices, want 1", len(devs))
+	}
+	if devs[0].Name != "/dev/mmcblk0" || devs[0].Type != "emmc" {
+		t.Fatalf("scanEmmcDevices()[0] = %+v, want Name=/dev/mmcblk0 Type=emmc", devs[0])
+	}
+
+	sm := &SmartManager{SmartDataMap: map[string]*smart.SmartData{}}
+	ok, err := sm.collectEmmcHealth(devs[0])
+	if err != nil || !ok {
+		t.Fatalf("collectEmmcHealth() = (ok=%v, err=%v), want (true,nil)", ok, err)
+	}
+	if len(sm.SmartDataMap) != 1 {
+		t.Fatalf("SmartDataMap len=%d, want 1", len(sm.SmartDataMap))
+	}
+	var got *smart.SmartData
+	for _, v := range sm.SmartDataMap {
+		got = v
+		break
+	}
+	if got == nil {
+		t.Fatalf("SmartDataMap value nil")
+	}
+	if got.DiskType != "emmc" || got.DiskName != "/dev/mmcblk0" {
+		t.Fatalf("disk fields = (type=%q name=%q), want (emmc,/dev/mmcblk0)", got.DiskType, got.DiskName)
+	}
+	if got.SmartStatus != "WARNING" {
+		t.Fatalf("SmartStatus=%q, want WARNING", got.SmartStatus)
+	}
+	if got.SerialNumber != "01234567" || got.ModelName == "" || got.Capacity == 0 {
+		t.Fatalf("identity fields = (model=%q serial=%q cap=%d), want non-empty model, serial 01234567, cap>0", got.ModelName, got.SerialNumber, got.Capacity)
+	}
+	if len(got.Attributes) < 3 {
+		t.Fatalf("attributes len=%d, want >= 3", len(got.Attributes))
+	}
+}

agent/emmc_stub.go

@@ -0,0 +1,14 @@
+//go:build !linux
+
+package agent
+
+// Non-Linux builds: eMMC health via sysfs is not available.
+
+func scanEmmcDevices() []*DeviceInfo {
+	return nil
+}
+
+func (sm *SmartManager) collectEmmcHealth(deviceInfo *DeviceInfo) (bool, error) {
+	return false, nil
+}
+

agent/gpu.go

@@ -9,6 +9,7 @@ import (
 	"maps"
 	"os/exec"
 	"regexp"
+	"runtime"
 	"strconv"
 	"strings"
 	"sync"
@@ -19,15 +20,14 @@ import (
 
 const (
 	// Commands
-	nvidiaSmiCmd  string = "nvidia-smi"
-	rocmSmiCmd    string = "rocm-smi"
-	amdgpuCmd     string = "amdgpu" // internal cmd for sysfs collection
-	tegraStatsCmd string = "tegrastats"
+	nvidiaSmiCmd    string = "nvidia-smi"
+	rocmSmiCmd      string = "rocm-smi"
+	tegraStatsCmd   string = "tegrastats"
+	nvtopCmd        string = "nvtop"
+	powermetricsCmd string = "powermetrics"
+	macmonCmd       string = "macmon"
+	noGPUFoundMsg   string = "no GPU found - see https://beszel.dev/guide/gpu"
 
-	// Polling intervals
-	nvidiaSmiInterval  string        = "4"    // in seconds
-	tegraStatsInterval string        = "3700" // in milliseconds
-	rocmSmiInterval    time.Duration = 4300 * time.Millisecond
 	// Command retry and timeout constants
 	retryWaitTime     time.Duration = 5 * time.Second
 	maxFailureRetries int           = 5
@@ -40,13 +40,7 @@ const (
 // GPUManager manages data collection for GPUs (either Nvidia or AMD)
 type GPUManager struct {
 	sync.Mutex
-	nvidiaSmi     bool
-	rocmSmi       bool
-	amdgpu        bool
-	tegrastats    bool
-	intelGpuStats bool
-	nvml          bool
-	GpuDataMap    map[string]*system.GPUData
+	GpuDataMap map[string]*system.GPUData
 	// lastAvgData stores the last calculated averages for each GPU
 	// Used when a collection happens before new data arrives (Count == 0)
 	lastAvgData map[string]system.GPUData
@@ -87,6 +81,58 @@ type gpuCollector struct {
 
 var errNoValidData = fmt.Errorf("no valid GPU data found") // Error for missing data
 
+// collectorSource identifies a selectable GPU collector in GPU_COLLECTOR.
+type collectorSource string
+
+const (
+	collectorSourceNVTop        collectorSource = collectorSource(nvtopCmd)
+	collectorSourceNVML         collectorSource = "nvml"
+	collectorSourceNvidiaSMI    collectorSource = collectorSource(nvidiaSmiCmd)
+	collectorSourceIntelGpuTop  collectorSource = collectorSource(intelGpuStatsCmd)
+	collectorSourceAmdSysfs     collectorSource = "amd_sysfs"
+	collectorSourceRocmSMI      collectorSource = collectorSource(rocmSmiCmd)
+	collectorSourceMacmon       collectorSource = collectorSource(macmonCmd)
+	collectorSourcePowermetrics collectorSource = collectorSource(powermetricsCmd)
+	collectorGroupNvidia        string          = "nvidia"
+	collectorGroupIntel         string          = "intel"
+	collectorGroupAmd           string          = "amd"
+	collectorGroupApple         string          = "apple"
+)
+
+func isValidCollectorSource(source collectorSource) bool {
+	switch source {
+	case collectorSourceNVTop,
+		collectorSourceNVML,
+		collectorSourceNvidiaSMI,
+		collectorSourceIntelGpuTop,
+		collectorSourceAmdSysfs,
+		collectorSourceRocmSMI,
+		collectorSourceMacmon,
+		collectorSourcePowermetrics:
+		return true
+	}
+	return false
+}
+
+// gpuCapabilities describes detected GPU tooling and sysfs support on the host.
+type gpuCapabilities struct {
+	hasNvidiaSmi    bool
+	hasRocmSmi      bool
+	hasAmdSysfs     bool
+	hasTegrastats   bool
+	hasIntelGpuTop  bool
+	hasNvtop        bool
+	hasMacmon       bool
+	hasPowermetrics bool
+}
+
+type collectorDefinition struct {
+	group              string
+	available          bool
+	start              func(onFailure func()) bool
+	deprecationWarning string
+}
+
 // starts and manages the ongoing collection of GPU data for the specified GPU management utility
 func (c *gpuCollector) start() {
 	for {
@@ -392,93 +438,284 @@ func (gm *GPUManager) storeSnapshot(id string, gpu *system.GPUData, cacheKey uin
 	gm.lastSnapshots[cacheKey][id] = snapshot
 }
 
-// detectGPUs checks for the presence of GPU management tools (nvidia-smi, rocm-smi, tegrastats)
-// in the system path. It sets the corresponding flags in the GPUManager struct if any of these
-// tools are found. If none of the tools are found, it returns an error indicating that no GPU
-// management tools are available.
-func (gm *GPUManager) detectGPUs() error {
+// discoverGpuCapabilities checks for available GPU tooling and sysfs support.
+// It only reports capability presence and does not apply policy decisions.
+func (gm *GPUManager) discoverGpuCapabilities() gpuCapabilities {
+	caps := gpuCapabilities{
+		hasAmdSysfs: gm.hasAmdSysfs(),
+	}
 	if _, err := exec.LookPath(nvidiaSmiCmd); err == nil {
-		gm.nvidiaSmi = true
+		caps.hasNvidiaSmi = true
 	}
 	if _, err := exec.LookPath(rocmSmiCmd); err == nil {
-		if val, _ := GetEnv("AMD_SYSFS"); val == "true" {
-			gm.amdgpu = true
-		} else {
-			gm.rocmSmi = true
-		}
-	} else if gm.hasAmdSysfs() {
-		gm.amdgpu = true
+		caps.hasRocmSmi = true
 	}
 	if _, err := exec.LookPath(tegraStatsCmd); err == nil {
-		gm.tegrastats = true
-		gm.nvidiaSmi = false
+		caps.hasTegrastats = true
 	}
 	if _, err := exec.LookPath(intelGpuStatsCmd); err == nil {
-		gm.intelGpuStats = true
+		caps.hasIntelGpuTop = true
 	}
-	if gm.nvidiaSmi || gm.rocmSmi || gm.amdgpu || gm.tegrastats || gm.intelGpuStats || gm.nvml {
-		return nil
+	if _, err := exec.LookPath(nvtopCmd); err == nil {
+		caps.hasNvtop = true
 	}
-	return fmt.Errorf("no GPU found - install nvidia-smi, rocm-smi, or intel_gpu_top")
+	if runtime.GOOS == "darwin" {
+		if _, err := exec.LookPath(macmonCmd); err == nil {
+			caps.hasMacmon = true
+		}
+		if _, err := exec.LookPath(powermetricsCmd); err == nil {
+			caps.hasPowermetrics = true
+		}
+	}
+	return caps
 }
 
-// startCollector starts the appropriate GPU data collector based on the command
-func (gm *GPUManager) startCollector(command string) {
-	collector := gpuCollector{
-		name:    command,
-		bufSize: 10 * 1024,
-	}
-	switch command {
-	case intelGpuStatsCmd:
-		go func() {
-			failures := 0
-			for {
-				if err := gm.collectIntelStats(); err != nil {
-					failures++
-					if failures > maxFailureRetries {
-						break
-					}
-					slog.Warn("Error collecting Intel GPU data; see https://beszel.dev/guide/gpu", "err", err)
-					time.Sleep(retryWaitTime)
-					continue
+func hasAnyGpuCollector(caps gpuCapabilities) bool {
+	return caps.hasNvidiaSmi || caps.hasRocmSmi || caps.hasAmdSysfs || caps.hasTegrastats || caps.hasIntelGpuTop || caps.hasNvtop || caps.hasMacmon || caps.hasPowermetrics
+}
+
+func (gm *GPUManager) startIntelCollector() {
+	go func() {
+		failures := 0
+		for {
+			if err := gm.collectIntelStats(); err != nil {
+				failures++
+				if failures > maxFailureRetries {
+					break
 				}
+				slog.Warn("Error collecting Intel GPU data; see https://beszel.dev/guide/gpu", "err", err)
+				time.Sleep(retryWaitTime)
+				continue
 			}
-		}()
-	case nvidiaSmiCmd:
-		collector.cmdArgs = []string{
-			"-l", nvidiaSmiInterval,
+		}
+	}()
+}
+
+func (gm *GPUManager) startNvidiaSmiCollector(intervalSeconds string) {
+	collector := gpuCollector{
+		name:    nvidiaSmiCmd,
+		bufSize: 10 * 1024,
+		cmdArgs: []string{
+			"-l", intervalSeconds,
 			"--query-gpu=index,name,temperature.gpu,memory.used,memory.total,utilization.gpu,power.draw",
 			"--format=csv,noheader,nounits",
-		}
-		collector.parse = gm.parseNvidiaData
-		go collector.start()
-	case tegraStatsCmd:
-		collector.cmdArgs = []string{"--interval", tegraStatsInterval}
-		collector.parse = gm.getJetsonParser()
-		go collector.start()
-	case amdgpuCmd:
-		go func() {
-			if err := gm.collectAmdStats(); err != nil {
-				slog.Warn("Error collecting AMD GPU data via sysfs", "err", err)
-			}
-		}()
-	case rocmSmiCmd:
-		collector.cmdArgs = []string{"--showid", "--showtemp", "--showuse", "--showpower", "--showproductname", "--showmeminfo", "vram", "--json"}
-		collector.parse = gm.parseAmdData
-		go func() {
-			failures := 0
-			for {
-				if err := collector.collect(); err != nil {
-					failures++
-					if failures > maxFailureRetries {
-						break
-					}
-					slog.Warn("Error collecting AMD GPU data via rocm-smi", "err", err)
-				}
-				time.Sleep(rocmSmiInterval)
-			}
-		}()
+		},
+		parse: gm.parseNvidiaData,
 	}
+	go collector.start()
+}
+
+func (gm *GPUManager) startTegraStatsCollector(intervalMilliseconds string) {
+	collector := gpuCollector{
+		name:    tegraStatsCmd,
+		bufSize: 10 * 1024,
+		cmdArgs: []string{"--interval", intervalMilliseconds},
+		parse:   gm.getJetsonParser(),
+	}
+	go collector.start()
+}
+
+func (gm *GPUManager) startRocmSmiCollector(pollInterval time.Duration) {
+	collector := gpuCollector{
+		name:    rocmSmiCmd,
+		bufSize: 10 * 1024,
+		cmdArgs: []string{"--showid", "--showtemp", "--showuse", "--showpower", "--showproductname", "--showmeminfo", "vram", "--json"},
+		parse:   gm.parseAmdData,
+	}
+	go func() {
+		failures := 0
+		for {
+			if err := collector.collect(); err != nil {
+				failures++
+				if failures > maxFailureRetries {
+					break
+				}
+				slog.Warn("Error collecting AMD GPU data via rocm-smi", "err", err)
+			}
+			time.Sleep(pollInterval)
+		}
+	}()
+}
+
+func (gm *GPUManager) collectorDefinitions(caps gpuCapabilities) map[collectorSource]collectorDefinition {
+	return map[collectorSource]collectorDefinition{
+		collectorSourceNVML: {
+			group:     collectorGroupNvidia,
+			available: caps.hasNvidiaSmi,
+			start: func(_ func()) bool {
+				return gm.startNvmlCollector()
+			},
+		},
+		collectorSourceNvidiaSMI: {
+			group:     collectorGroupNvidia,
+			available: caps.hasNvidiaSmi,
+			start: func(_ func()) bool {
+				gm.startNvidiaSmiCollector("4") // seconds
+				return true
+			},
+		},
+		collectorSourceIntelGpuTop: {
+			group:     collectorGroupIntel,
+			available: caps.hasIntelGpuTop,
+			start: func(_ func()) bool {
+				gm.startIntelCollector()
+				return true
+			},
+		},
+		collectorSourceAmdSysfs: {
+			group:     collectorGroupAmd,
+			available: caps.hasAmdSysfs,
+			start: func(_ func()) bool {
+				return gm.startAmdSysfsCollector()
+			},
+		},
+		collectorSourceRocmSMI: {
+			group:              collectorGroupAmd,
+			available:          caps.hasRocmSmi,
+			deprecationWarning: "rocm-smi is deprecated and may be removed in a future release",
+			start: func(_ func()) bool {
+				gm.startRocmSmiCollector(4300 * time.Millisecond)
+				return true
+			},
+		},
+		collectorSourceNVTop: {
+			available: caps.hasNvtop,
+			start: func(onFailure func()) bool {
+				gm.startNvtopCollector("30", onFailure) // tens of milliseconds
+				return true
+			},
+		},
+		collectorSourceMacmon: {
+			group:     collectorGroupApple,
+			available: caps.hasMacmon,
+			start: func(_ func()) bool {
+				gm.startMacmonCollector()
+				return true
+			},
+		},
+		collectorSourcePowermetrics: {
+			group:     collectorGroupApple,
+			available: caps.hasPowermetrics,
+			start: func(_ func()) bool {
+				gm.startPowermetricsCollector()
+				return true
+			},
+		},
+	}
+}
+
+// parseCollectorPriority parses GPU_COLLECTOR and returns valid ordered entries.
+func parseCollectorPriority(value string) []collectorSource {
+	parts := strings.Split(value, ",")
+	priorities := make([]collectorSource, 0, len(parts))
+	for _, raw := range parts {
+		name := collectorSource(strings.TrimSpace(strings.ToLower(raw)))
+		if !isValidCollectorSource(name) {
+			if name != "" {
+				slog.Warn("Ignoring unknown GPU collector", "collector", name)
+			}
+			continue
+		}
+		priorities = append(priorities, name)
+	}
+	return priorities
+}
+
+// startNvmlCollector initializes NVML and starts its polling loop.
+func (gm *GPUManager) startNvmlCollector() bool {
+	collector := &nvmlCollector{gm: gm}
+	if err := collector.init(); err != nil {
+		slog.Warn("Failed to initialize NVML", "err", err)
+		return false
+	}
+	go collector.start()
+	return true
+}
+
+// startAmdSysfsCollector starts AMD GPU collection via sysfs.
+func (gm *GPUManager) startAmdSysfsCollector() bool {
+	go func() {
+		if err := gm.collectAmdStats(); err != nil {
+			slog.Warn("Error collecting AMD GPU data via sysfs", "err", err)
+		}
+	}()
+	return true
+}
+
+// startCollectorsByPriority starts collectors in order with one source per vendor group.
+func (gm *GPUManager) startCollectorsByPriority(priorities []collectorSource, caps gpuCapabilities) int {
+	definitions := gm.collectorDefinitions(caps)
+	selectedGroups := make(map[string]bool, 3)
+	started := 0
+	for i, source := range priorities {
+		definition, ok := definitions[source]
+		if !ok || !definition.available {
+			continue
+		}
+		// nvtop is not a vendor-specific collector, so should only be used if no other collectors are selected or it is first in GPU_COLLECTOR.
+		if source == collectorSourceNVTop {
+			if len(selectedGroups) > 0 {
+				slog.Warn("Skipping nvtop because other collectors are selected")
+				continue
+			}
+			// if nvtop fails, fall back to remaining collectors.
+			remaining := append([]collectorSource(nil), priorities[i+1:]...)
+			if definition.start(func() {
+				gm.startCollectorsByPriority(remaining, caps)
+			}) {
+				started++
+				return started
+			}
+		}
+		group := definition.group
+		if group == "" || selectedGroups[group] {
+			continue
+		}
+		if definition.deprecationWarning != "" {
+			slog.Warn(definition.deprecationWarning)
+		}
+		if definition.start(nil) {
+			selectedGroups[group] = true
+			started++
+		}
+	}
+	return started
+}
+
+// resolveLegacyCollectorPriority builds the default collector order when GPU_COLLECTOR is unset.
+func (gm *GPUManager) resolveLegacyCollectorPriority(caps gpuCapabilities) []collectorSource {
+	priorities := make([]collectorSource, 0, 4)
+
+	if caps.hasNvidiaSmi && !caps.hasTegrastats {
+		if nvml, _ := GetEnv("NVML"); nvml == "true" {
+			priorities = append(priorities, collectorSourceNVML, collectorSourceNvidiaSMI)
+		} else {
+			priorities = append(priorities, collectorSourceNvidiaSMI)
+		}
+	}
+
+	if caps.hasRocmSmi {
+		if val, _ := GetEnv("AMD_SYSFS"); val == "true" {
+			priorities = append(priorities, collectorSourceAmdSysfs)
+		} else {
+			priorities = append(priorities, collectorSourceRocmSMI)
+		}
+	} else if caps.hasAmdSysfs {
+		priorities = append(priorities, collectorSourceAmdSysfs)
+	}
+
+	if caps.hasIntelGpuTop {
+		priorities = append(priorities, collectorSourceIntelGpuTop)
+	}
+
+	// Apple collectors are currently opt-in only.
+	// Enable them with GPU_COLLECTOR=macmon or GPU_COLLECTOR=powermetrics.
+
+	// Keep nvtop as a last resort only when no vendor collector exists.
+	if len(priorities) == 0 && caps.hasNvtop {
+		priorities = append(priorities, collectorSourceNVTop)
+	}
+	return priorities
 }
 
 // NewGPUManager creates and initializes a new GPUManager
@@ -487,38 +724,30 @@ func NewGPUManager() (*GPUManager, error) {
 		return nil, nil
 	}
 	var gm GPUManager
-	if err := gm.detectGPUs(); err != nil {
-		return nil, err
+	caps := gm.discoverGpuCapabilities()
+	if !hasAnyGpuCollector(caps) {
+		return nil, fmt.Errorf(noGPUFoundMsg)
 	}
 	gm.GpuDataMap = make(map[string]*system.GPUData)
 
-	if gm.nvidiaSmi {
-		if nvml, _ := GetEnv("NVML"); nvml == "true" {
-			gm.nvml = true
-			gm.nvidiaSmi = false
-			collector := &nvmlCollector{gm: &gm}
-			if err := collector.init(); err == nil {
-				go collector.start()
-			} else {
-				slog.Warn("Failed to initialize NVML, falling back to nvidia-smi", "err", err)
-				gm.nvidiaSmi = true
-				gm.startCollector(nvidiaSmiCmd)
-			}
-		} else {
-			gm.startCollector(nvidiaSmiCmd)
+	// Jetson devices should always use tegrastats (ignore GPU_COLLECTOR).
+	if caps.hasTegrastats {
+		gm.startTegraStatsCollector("3700")
+		return &gm, nil
+	}
+
+	// if GPU_COLLECTOR is set, start user-defined collectors.
+	if collectorConfig, ok := GetEnv("GPU_COLLECTOR"); ok && strings.TrimSpace(collectorConfig) != "" {
+		priorities := parseCollectorPriority(collectorConfig)
+		if gm.startCollectorsByPriority(priorities, caps) == 0 {
+			return nil, fmt.Errorf("no configured GPU collectors are available")
 		}
+		return &gm, nil
 	}
-	if gm.rocmSmi {
-		gm.startCollector(rocmSmiCmd)
-	}
-	if gm.amdgpu {
-		gm.startCollector(amdgpuCmd)
-	}
-	if gm.tegrastats {
-		gm.startCollector(tegraStatsCmd)
-	}
-	if gm.intelGpuStats {
-		gm.startCollector(intelGpuStatsCmd)
+
+	// auto-detect and start collectors when GPU_COLLECTOR is unset.
+	if gm.startCollectorsByPriority(gm.resolveLegacyCollectorPriority(caps), caps) == 0 {
+		return nil, fmt.Errorf(noGPUFoundMsg)
 	}
 
 	return &gm, nil

agent/gpu_amd_linux.go

@@ -3,6 +3,7 @@
 package agent
 
 import (
+	"bufio"
 	"fmt"
 	"log/slog"
 	"os"
@@ -15,6 +16,15 @@ import (
 	"github.com/henrygd/beszel/internal/entities/system"
 )
 
+var amdgpuNameCache = struct {
+	sync.RWMutex
+	hits   map[string]string
+	misses map[string]struct{}
+}{
+	hits:   make(map[string]string),
+	misses: make(map[string]struct{}),
+}
+
 // hasAmdSysfs returns true if any AMD GPU sysfs nodes are found
 func (gm *GPUManager) hasAmdSysfs() bool {
 	cards, err := filepath.Glob("/sys/class/drm/card*/device/vendor")
@@ -32,6 +42,7 @@ func (gm *GPUManager) hasAmdSysfs() bool {
 
 // collectAmdStats collects AMD GPU metrics directly from sysfs to avoid the overhead of rocm-smi
 func (gm *GPUManager) collectAmdStats() error {
+	sysfsPollInterval := 3000 * time.Millisecond
 	cards, err := filepath.Glob("/sys/class/drm/card*")
 	if err != nil {
 		return err
@@ -70,10 +81,11 @@ func (gm *GPUManager) collectAmdStats() error {
 			continue
 		}
 		failures = 0
-		time.Sleep(rocmSmiInterval)
+		time.Sleep(sysfsPollInterval)
 	}
 }
 
+// isAmdGpu checks whether a DRM card path belongs to AMD vendor ID 0x1002.
 func isAmdGpu(cardPath string) bool {
 	vendorPath := filepath.Join(cardPath, "device/vendor")
 	vendor, err := os.ReadFile(vendorPath)
@@ -91,8 +103,17 @@ func (gm *GPUManager) updateAmdGpuData(cardPath string) bool {
 
 	// Read all sysfs values first (no lock needed - these can be slow)
 	usage, usageErr := readSysfsFloat(filepath.Join(devicePath, "gpu_busy_percent"))
-	memUsed, memUsedErr := readSysfsFloat(filepath.Join(devicePath, "mem_info_vram_used"))
-	memTotal, _ := readSysfsFloat(filepath.Join(devicePath, "mem_info_vram_total"))
+	vramUsed, memUsedErr := readSysfsFloat(filepath.Join(devicePath, "mem_info_vram_used"))
+	vramTotal, _ := readSysfsFloat(filepath.Join(devicePath, "mem_info_vram_total"))
+	memUsed := vramUsed
+	memTotal := vramTotal
+	// if gtt is present, add it to the memory used and total (https://github.com/henrygd/beszel/issues/1569#issuecomment-3837640484)
+	if gttUsed, err := readSysfsFloat(filepath.Join(devicePath, "mem_info_gtt_used")); err == nil && gttUsed > 0 {
+		if gttTotal, err := readSysfsFloat(filepath.Join(devicePath, "mem_info_gtt_total")); err == nil {
+			memUsed += gttUsed
+			memTotal += gttTotal
+		}
+	}
 
 	var temp, power float64
 	hwmons, _ := filepath.Glob(filepath.Join(devicePath, "hwmon/hwmon*"))
@@ -133,6 +154,7 @@ func (gm *GPUManager) updateAmdGpuData(cardPath string) bool {
 	return true
 }
 
+// readSysfsFloat reads and parses a numeric value from a sysfs file.
 func readSysfsFloat(path string) (float64, error) {
 	val, err := os.ReadFile(path)
 	if err != nil {
@@ -141,6 +163,110 @@ func readSysfsFloat(path string) (float64, error) {
 	return strconv.ParseFloat(strings.TrimSpace(string(val)), 64)
 }
 
+// normalizeHexID normalizes hex IDs by trimming spaces, lowercasing, and dropping 0x.
+func normalizeHexID(id string) string {
+	return strings.TrimPrefix(strings.ToLower(strings.TrimSpace(id)), "0x")
+}
+
+// cacheKeyForAmdgpu builds the cache key for a device and optional revision.
+func cacheKeyForAmdgpu(deviceID, revisionID string) string {
+	if revisionID != "" {
+		return deviceID + ":" + revisionID
+	}
+	return deviceID
+}
+
+// lookupAmdgpuNameInFile resolves an AMDGPU name from amdgpu.ids by device/revision.
+func lookupAmdgpuNameInFile(deviceID, revisionID, filePath string) (name string, exact bool, found bool) {
+	file, err := os.Open(filePath)
+	if err != nil {
+		return "", false, false
+	}
+	defer file.Close()
+
+	var byDevice string
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" || strings.HasPrefix(line, "#") {
+			continue
+		}
+		parts := strings.SplitN(line, ",", 3)
+		if len(parts) != 3 {
+			continue
+		}
+
+		dev := normalizeHexID(parts[0])
+		rev := normalizeHexID(parts[1])
+		productName := strings.TrimSpace(parts[2])
+		if dev == "" || productName == "" || dev != deviceID {
+			continue
+		}
+		if byDevice == "" {
+			byDevice = productName
+		}
+		if revisionID != "" && rev == revisionID {
+			return productName, true, true
+		}
+	}
+	if byDevice != "" {
+		return byDevice, false, true
+	}
+	return "", false, false
+}
+
+// getCachedAmdgpuName returns cached hit/miss status for the given device/revision.
+func getCachedAmdgpuName(deviceID, revisionID string) (name string, found bool, done bool) {
+	// Build the list of cache keys to check. We always look up the exact device+revision key.
+	// When revisionID is set, we also look up deviceID alone, since the cache may store a
+	// device-only fallback when we couldn't resolve the exact revision.
+	keys := []string{cacheKeyForAmdgpu(deviceID, revisionID)}
+	if revisionID != "" {
+		keys = append(keys, deviceID)
+	}
+
+	knownMisses := 0
+	amdgpuNameCache.RLock()
+	defer amdgpuNameCache.RUnlock()
+	for _, key := range keys {
+		if name, ok := amdgpuNameCache.hits[key]; ok {
+			return name, true, true
+		}
+		if _, ok := amdgpuNameCache.misses[key]; ok {
+			knownMisses++
+		}
+	}
+	// done=true means "don't bother doing slow lookup": we either found a name (above) or
+	// every key we checked was already a known miss, so we've tried before and failed.
+	return "", false, knownMisses == len(keys)
+}
+
+// normalizeAmdgpuName trims standard suffixes from AMDGPU product names.
+func normalizeAmdgpuName(name string) string {
+	return strings.TrimSuffix(strings.TrimSpace(name), " Graphics")
+}
+
+// cacheAmdgpuName stores a resolved AMDGPU name in the lookup cache.
+func cacheAmdgpuName(deviceID, revisionID, name string, exact bool) {
+	name = normalizeAmdgpuName(name)
+	amdgpuNameCache.Lock()
+	defer amdgpuNameCache.Unlock()
+	if exact && revisionID != "" {
+		amdgpuNameCache.hits[cacheKeyForAmdgpu(deviceID, revisionID)] = name
+	}
+	amdgpuNameCache.hits[deviceID] = name
+}
+
+// cacheMissingAmdgpuName records unresolved device/revision lookups.
+func cacheMissingAmdgpuName(deviceID, revisionID string) {
+	amdgpuNameCache.Lock()
+	defer amdgpuNameCache.Unlock()
+	amdgpuNameCache.misses[deviceID] = struct{}{}
+	if revisionID != "" {
+		amdgpuNameCache.misses[cacheKeyForAmdgpu(deviceID, revisionID)] = struct{}{}
+	}
+}
+
 // getAmdGpuName attempts to get a descriptive GPU name.
 // First tries product_name (rarely available), then looks up the PCI device ID.
 // Falls back to showing the raw device ID if not found in the lookup table.
@@ -152,33 +278,24 @@ func getAmdGpuName(devicePath string) string {
 
 	// Read PCI device ID and look it up
 	if deviceID, err := os.ReadFile(filepath.Join(devicePath, "device")); err == nil {
-		id := strings.TrimPrefix(strings.ToLower(strings.TrimSpace(string(deviceID))), "0x")
-		if name, ok := getRadeonNames()[id]; ok {
-			return fmt.Sprintf("Radeon %s", name)
+		id := normalizeHexID(string(deviceID))
+		revision := ""
+		if revBytes, revErr := os.ReadFile(filepath.Join(devicePath, "revision")); revErr == nil {
+			revision = normalizeHexID(string(revBytes))
 		}
+
+		if name, found, done := getCachedAmdgpuName(id, revision); found {
+			return name
+		} else if !done {
+			if name, exact, ok := lookupAmdgpuNameInFile(id, revision, "/usr/share/libdrm/amdgpu.ids"); ok {
+				cacheAmdgpuName(id, revision, name, exact)
+				return normalizeAmdgpuName(name)
+			}
+			cacheMissingAmdgpuName(id, revision)
+		}
+
 		return fmt.Sprintf("AMD GPU (%s)", id)
 	}
 
 	return "AMD GPU"
 }
-
-// getRadeonNames returns the AMD GPU name lookup table
-// Device IDs from https://pci-ids.ucw.cz/read/PC/1002
-var getRadeonNames = sync.OnceValue(func() map[string]string {
-	return map[string]string{
-		"7550": "RX 9070",
-		"7590": "RX 9060 XT",
-		"7551": "AI PRO R9700",
-
-		"744c": "RX 7900",
-
-		"1681": "680M",
-
-		"7448": "PRO W7900",
-		"745e": "PRO W7800",
-		"7470": "PRO W7700",
-		"73e3": "PRO W6600",
-		"7422": "PRO W6400",
-		"7341": "PRO W5500",
-	}
-})

agent/gpu_amd_linux_test.go

@@ -0,0 +1,264 @@
+//go:build linux
+
+package agent
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestNormalizeHexID(t *testing.T) {
+	tests := []struct {
+		in   string
+		want string
+	}{
+		{"0x1002", "1002"},
+		{"C2", "c2"},
+		{"  15BF  ", "15bf"},
+		{"0x15bf", "15bf"},
+		{"", ""},
+	}
+	for _, tt := range tests {
+		subName := tt.in
+		if subName == "" {
+			subName = "empty_string"
+		}
+		t.Run(subName, func(t *testing.T) {
+			got := normalizeHexID(tt.in)
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestCacheKeyForAmdgpu(t *testing.T) {
+	tests := []struct {
+		deviceID   string
+		revisionID string
+		want       string
+	}{
+		{"1114", "c2", "1114:c2"},
+		{"15bf", "", "15bf"},
+		{"1506", "c1", "1506:c1"},
+	}
+	for _, tt := range tests {
+		got := cacheKeyForAmdgpu(tt.deviceID, tt.revisionID)
+		assert.Equal(t, tt.want, got)
+	}
+}
+
+func TestReadSysfsFloat(t *testing.T) {
+	dir := t.TempDir()
+
+	validPath := filepath.Join(dir, "val")
+	require.NoError(t, os.WriteFile(validPath, []byte("  42.5  \n"), 0o644))
+	got, err := readSysfsFloat(validPath)
+	require.NoError(t, err)
+	assert.Equal(t, 42.5, got)
+
+	// Integer and scientific
+	sciPath := filepath.Join(dir, "sci")
+	require.NoError(t, os.WriteFile(sciPath, []byte("1e2"), 0o644))
+	got, err = readSysfsFloat(sciPath)
+	require.NoError(t, err)
+	assert.Equal(t, 100.0, got)
+
+	// Missing file
+	_, err = readSysfsFloat(filepath.Join(dir, "missing"))
+	require.Error(t, err)
+
+	// Invalid content
+	badPath := filepath.Join(dir, "bad")
+	require.NoError(t, os.WriteFile(badPath, []byte("not a number"), 0o644))
+	_, err = readSysfsFloat(badPath)
+	require.Error(t, err)
+}
+
+func TestIsAmdGpu(t *testing.T) {
+	dir := t.TempDir()
+	deviceDir := filepath.Join(dir, "device")
+	require.NoError(t, os.MkdirAll(deviceDir, 0o755))
+
+	// AMD vendor 0x1002 -> true
+	require.NoError(t, os.WriteFile(filepath.Join(deviceDir, "vendor"), []byte("0x1002\n"), 0o644))
+	assert.True(t, isAmdGpu(dir), "vendor 0x1002 should be AMD")
+
+	// Non-AMD vendor -> false
+	require.NoError(t, os.WriteFile(filepath.Join(deviceDir, "vendor"), []byte("0x10de\n"), 0o644))
+	assert.False(t, isAmdGpu(dir), "vendor 0x10de should not be AMD")
+
+	// Missing vendor file -> false
+	require.NoError(t, os.Remove(filepath.Join(deviceDir, "vendor")))
+	assert.False(t, isAmdGpu(dir), "missing vendor file should be false")
+}
+
+func TestAmdgpuNameCacheRoundTrip(t *testing.T) {
+	// Cache a name and retrieve it (unique key to avoid affecting other tests)
+	deviceID, revisionID := "cachedev99", "00"
+	cacheAmdgpuName(deviceID, revisionID, "AMD Test GPU 99 Graphics", true)
+
+	name, found, done := getCachedAmdgpuName(deviceID, revisionID)
+	assert.True(t, found)
+	assert.True(t, done)
+	assert.Equal(t, "AMD Test GPU 99", name)
+
+	// Device-only key also stored
+	name2, found2, _ := getCachedAmdgpuName(deviceID, "")
+	assert.True(t, found2)
+	assert.Equal(t, "AMD Test GPU 99", name2)
+
+	// Cache a miss
+	cacheMissingAmdgpuName("missedev99", "ab")
+	_, found3, done3 := getCachedAmdgpuName("missedev99", "ab")
+	assert.False(t, found3)
+	assert.True(t, done3, "done should be true so caller skips file lookup")
+}
+
+func TestUpdateAmdGpuDataWithFakeSysfs(t *testing.T) {
+	tests := []struct {
+		name            string
+		writeGTT        bool
+		wantMemoryUsed  float64
+		wantMemoryTotal float64
+	}{
+		{
+			name:            "sums vram and gtt when gtt is present",
+			writeGTT:        true,
+			wantMemoryUsed:  bytesToMegabytes(1073741824 + 536870912),
+			wantMemoryTotal: bytesToMegabytes(2147483648 + 4294967296),
+		},
+		{
+			name:            "falls back to vram when gtt is missing",
+			writeGTT:        false,
+			wantMemoryUsed:  bytesToMegabytes(1073741824),
+			wantMemoryTotal: bytesToMegabytes(2147483648),
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			cardPath := filepath.Join(dir, "card0")
+			devicePath := filepath.Join(cardPath, "device")
+			hwmonPath := filepath.Join(devicePath, "hwmon", "hwmon0")
+			require.NoError(t, os.MkdirAll(hwmonPath, 0o755))
+
+			write := func(name, content string) {
+				require.NoError(t, os.WriteFile(filepath.Join(devicePath, name), []byte(content), 0o644))
+			}
+			write("vendor", "0x1002")
+			write("device", "0x1506")
+			write("revision", "0xc1")
+			write("gpu_busy_percent", "25")
+			write("mem_info_vram_used", "1073741824")
+			write("mem_info_vram_total", "2147483648")
+			if tt.writeGTT {
+				write("mem_info_gtt_used", "536870912")
+				write("mem_info_gtt_total", "4294967296")
+			}
+			require.NoError(t, os.WriteFile(filepath.Join(hwmonPath, "temp1_input"), []byte("45000"), 0o644))
+			require.NoError(t, os.WriteFile(filepath.Join(hwmonPath, "power1_input"), []byte("20000000"), 0o644))
+
+			// Pre-cache name so getAmdGpuName returns a known value (it uses system amdgpu.ids path)
+			cacheAmdgpuName("1506", "c1", "AMD Radeon 610M Graphics", true)
+
+			gm := &GPUManager{GpuDataMap: make(map[string]*system.GPUData)}
+			ok := gm.updateAmdGpuData(cardPath)
+			require.True(t, ok)
+
+			gpu, ok := gm.GpuDataMap["card0"]
+			require.True(t, ok)
+			assert.Equal(t, "AMD Radeon 610M", gpu.Name)
+			assert.Equal(t, 25.0, gpu.Usage)
+			assert.Equal(t, tt.wantMemoryUsed, gpu.MemoryUsed)
+			assert.Equal(t, tt.wantMemoryTotal, gpu.MemoryTotal)
+			assert.Equal(t, 45.0, gpu.Temperature)
+			assert.Equal(t, 20.0, gpu.Power)
+			assert.Equal(t, 1.0, gpu.Count)
+		})
+	}
+}
+
+func TestLookupAmdgpuNameInFile(t *testing.T) {
+	idsPath := filepath.Join("test-data", "amdgpu.ids")
+
+	tests := []struct {
+		name       string
+		deviceID   string
+		revisionID string
+		wantName   string
+		wantExact  bool
+		wantFound  bool
+	}{
+		{
+			name:       "exact device and revision match",
+			deviceID:   "1114",
+			revisionID: "c2",
+			wantName:   "AMD Radeon 860M Graphics",
+			wantExact:  true,
+			wantFound:  true,
+		},
+		{
+			name:       "exact match 15BF revision 01 returns 760M",
+			deviceID:   "15bf",
+			revisionID: "01",
+			wantName:   "AMD Radeon 760M Graphics",
+			wantExact:  true,
+			wantFound:  true,
+		},
+		{
+			name:       "exact match 15BF revision 00 returns 780M",
+			deviceID:   "15bf",
+			revisionID: "00",
+			wantName:   "AMD Radeon 780M Graphics",
+			wantExact:  true,
+			wantFound:  true,
+		},
+		{
+			name:       "device-only match returns first entry for device",
+			deviceID:   "1506",
+			revisionID: "",
+			wantName:   "AMD Radeon 610M",
+			wantExact:  false,
+			wantFound:  true,
+		},
+		{
+			name:       "unknown device not found",
+			deviceID:   "dead",
+			revisionID: "00",
+			wantName:   "",
+			wantExact:  false,
+			wantFound:  false,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			gotName, gotExact, gotFound := lookupAmdgpuNameInFile(tt.deviceID, tt.revisionID, idsPath)
+			assert.Equal(t, tt.wantName, gotName, "name")
+			assert.Equal(t, tt.wantExact, gotExact, "exact")
+			assert.Equal(t, tt.wantFound, gotFound, "found")
+		})
+	}
+}
+
+func TestGetAmdGpuNameFromIdsFile(t *testing.T) {
+	// Test that getAmdGpuName resolves a name when we can't inject the ids path.
+	// We only verify behavior when product_name is missing and device/revision
+	// would be read from sysfs; the actual lookup uses /usr/share/libdrm/amdgpu.ids.
+	// So this test focuses on normalizeAmdgpuName and that lookupAmdgpuNameInFile
+	// returns the expected name for our test-data file.
+	idsPath := filepath.Join("test-data", "amdgpu.ids")
+	name, exact, found := lookupAmdgpuNameInFile("1435", "ae", idsPath)
+	require.True(t, found)
+	require.True(t, exact)
+	assert.Equal(t, "AMD Custom GPU 0932", name)
+	assert.Equal(t, "AMD Custom GPU 0932", normalizeAmdgpuName(name))
+
+	// " Graphics" suffix is trimmed by normalizeAmdgpuName
+	name2 := "AMD Radeon 860M Graphics"
+	assert.Equal(t, "AMD Radeon 860M", normalizeAmdgpuName(name2))
+}

agent/gpu_apple_unsupported.go

@@ -0,0 +1,9 @@
+//go:build !darwin
+
+package agent
+
+// startPowermetricsCollector is a no-op on non-darwin platforms; the real implementation is in gpu_darwin.go.
+func (gm *GPUManager) startPowermetricsCollector() {}
+
+// startMacmonCollector is a no-op on non-darwin platforms; the real implementation is in gpu_darwin.go.
+func (gm *GPUManager) startMacmonCollector() {}

agent/gpu_darwin.go

@@ -0,0 +1,252 @@
+//go:build darwin
+
+package agent
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"io"
+	"log/slog"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+)
+
+const (
+	// powermetricsSampleIntervalMs is the sampling interval passed to powermetrics (-i).
+	powermetricsSampleIntervalMs = 500
+	// powermetricsPollInterval is how often we run powermetrics to collect a new sample.
+	powermetricsPollInterval = 2 * time.Second
+	// macmonIntervalMs is the sampling interval passed to macmon pipe (-i), in milliseconds.
+	macmonIntervalMs = 2500
+)
+
+const appleGPUID = "0"
+
+// startPowermetricsCollector runs powermetrics --samplers gpu_power in a loop and updates
+// GPU usage and power. Requires root (sudo) on macOS. A single logical GPU is reported as id "0".
+func (gm *GPUManager) startPowermetricsCollector() {
+	// Ensure single GPU entry for Apple GPU
+	if _, ok := gm.GpuDataMap[appleGPUID]; !ok {
+		gm.GpuDataMap[appleGPUID] = &system.GPUData{Name: "Apple GPU"}
+	}
+
+	go func() {
+		failures := 0
+		for {
+			if err := gm.collectPowermetrics(); err != nil {
+				failures++
+				if failures > maxFailureRetries {
+					slog.Warn("powermetrics GPU collector failed repeatedly, stopping", "err", err)
+					break
+				}
+				slog.Warn("Error collecting macOS GPU data via powermetrics (may require sudo)", "err", err)
+				time.Sleep(retryWaitTime)
+				continue
+			}
+			failures = 0
+			time.Sleep(powermetricsPollInterval)
+		}
+	}()
+}
+
+// collectPowermetrics runs powermetrics once and parses GPU usage and power from its output.
+func (gm *GPUManager) collectPowermetrics() error {
+	interval := strconv.Itoa(powermetricsSampleIntervalMs)
+	cmd := exec.Command(powermetricsCmd, "--samplers", "gpu_power", "-i", interval, "-n", "1")
+	cmd.Stderr = nil
+	out, err := cmd.Output()
+	if err != nil {
+		return err
+	}
+	if !gm.parsePowermetricsData(out) {
+		return errNoValidData
+	}
+	return nil
+}
+
+// parsePowermetricsData parses powermetrics gpu_power output and updates GpuDataMap["0"].
+// Example output:
+//
+//	**** GPU usage ****
+//	GPU HW active frequency: 444 MHz
+//	GPU HW active residency:   0.97% (444 MHz: .97% ...
+//	GPU idle residency:  99.03%
+//	GPU Power: 4 mW
+func (gm *GPUManager) parsePowermetricsData(output []byte) bool {
+	var idleResidency, powerMW float64
+	var gotIdle, gotPower bool
+
+	scanner := bufio.NewScanner(bytes.NewReader(output))
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if strings.HasPrefix(line, "GPU idle residency:") {
+			// "GPU idle residency:  99.03%"
+			fields := strings.Fields(strings.TrimPrefix(line, "GPU idle residency:"))
+			if len(fields) >= 1 {
+				pct := strings.TrimSuffix(fields[0], "%")
+				if v, err := strconv.ParseFloat(pct, 64); err == nil {
+					idleResidency = v
+					gotIdle = true
+				}
+			}
+		} else if strings.HasPrefix(line, "GPU Power:") {
+			// "GPU Power: 4 mW"
+			fields := strings.Fields(strings.TrimPrefix(line, "GPU Power:"))
+			if len(fields) >= 1 {
+				if v, err := strconv.ParseFloat(fields[0], 64); err == nil {
+					powerMW = v
+					gotPower = true
+				}
+			}
+		}
+	}
+	if err := scanner.Err(); err != nil {
+		return false
+	}
+	if !gotIdle && !gotPower {
+		return false
+	}
+
+	gm.Lock()
+	defer gm.Unlock()
+
+	if _, ok := gm.GpuDataMap[appleGPUID]; !ok {
+		gm.GpuDataMap[appleGPUID] = &system.GPUData{Name: "Apple GPU"}
+	}
+	gpu := gm.GpuDataMap[appleGPUID]
+
+	if gotIdle {
+		// Usage = 100 - idle residency (e.g. 100 - 99.03 = 0.97%)
+		gpu.Usage += 100 - idleResidency
+	}
+	if gotPower {
+		// mW -> W
+		gpu.Power += powerMW / milliwattsInAWatt
+	}
+	gpu.Count++
+	return true
+}
+
+// startMacmonCollector runs `macmon pipe` in a loop and parses one JSON object per line.
+// This collector does not require sudo. A single logical GPU is reported as id "0".
+func (gm *GPUManager) startMacmonCollector() {
+	if _, ok := gm.GpuDataMap[appleGPUID]; !ok {
+		gm.GpuDataMap[appleGPUID] = &system.GPUData{Name: "Apple GPU"}
+	}
+
+	go func() {
+		failures := 0
+		for {
+			if err := gm.collectMacmonPipe(); err != nil {
+				failures++
+				if failures > maxFailureRetries {
+					slog.Warn("macmon GPU collector failed repeatedly, stopping", "err", err)
+					break
+				}
+				slog.Warn("Error collecting macOS GPU data via macmon", "err", err)
+				time.Sleep(retryWaitTime)
+				continue
+			}
+			failures = 0
+			// `macmon pipe` is long-running; if it returns, wait a bit before restarting.
+			time.Sleep(retryWaitTime)
+		}
+	}()
+}
+
+type macmonTemp struct {
+	GPUTempAvg float64 `json:"gpu_temp_avg"`
+}
+
+type macmonSample struct {
+	GPUPower    float64    `json:"gpu_power"`     // watts (macmon reports fractional values)
+	GPURAMPower float64    `json:"gpu_ram_power"` // watts
+	GPUUsage    []float64  `json:"gpu_usage"`     // [freq_mhz, usage] where usage is typically 0..1
+	Temp        macmonTemp `json:"temp"`
+}
+
+func (gm *GPUManager) collectMacmonPipe() (err error) {
+	cmd := exec.Command(macmonCmd, "pipe", "-i", strconv.Itoa(macmonIntervalMs))
+	// Avoid blocking if macmon writes to stderr.
+	cmd.Stderr = io.Discard
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return err
+	}
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+
+	// Ensure we always reap the child to avoid zombies on any return path and
+	// propagate a non-zero exit code if no other error was set.
+	defer func() {
+		_ = stdout.Close()
+		if cmd.ProcessState == nil || !cmd.ProcessState.Exited() {
+			_ = cmd.Process.Kill()
+		}
+		if waitErr := cmd.Wait(); err == nil && waitErr != nil {
+			err = waitErr
+		}
+	}()
+
+	scanner := bufio.NewScanner(stdout)
+	var hadSample bool
+	for scanner.Scan() {
+		line := bytes.TrimSpace(scanner.Bytes())
+		if len(line) == 0 {
+			continue
+		}
+		if gm.parseMacmonLine(line) {
+			hadSample = true
+		}
+	}
+	if scanErr := scanner.Err(); scanErr != nil {
+		return scanErr
+	}
+	if !hadSample {
+		return errNoValidData
+	}
+	return nil
+}
+
+// parseMacmonLine parses a single macmon JSON line and updates Apple GPU metrics.
+func (gm *GPUManager) parseMacmonLine(line []byte) bool {
+	var sample macmonSample
+	if err := json.Unmarshal(line, &sample); err != nil {
+		return false
+	}
+
+	usage := 0.0
+	if len(sample.GPUUsage) >= 2 {
+		usage = sample.GPUUsage[1]
+		// Heuristic: macmon typically reports 0..1; convert to percentage.
+		if usage <= 1.0 {
+			usage *= 100
+		}
+	}
+
+	// Consider the line valid if it contains at least one GPU metric.
+	if usage == 0 && sample.GPUPower == 0 && sample.Temp.GPUTempAvg == 0 {
+		return false
+	}
+
+	gm.Lock()
+	defer gm.Unlock()
+
+	gpu, ok := gm.GpuDataMap[appleGPUID]
+	if !ok {
+		gpu = &system.GPUData{Name: "Apple GPU"}
+		gm.GpuDataMap[appleGPUID] = gpu
+	}
+	gpu.Temperature = sample.Temp.GPUTempAvg
+	gpu.Usage += usage
+	// macmon reports power in watts; include VRAM power if present.
+	gpu.Power += sample.GPUPower + sample.GPURAMPower
+	gpu.Count++
+	return true
+}

agent/gpu_darwin_test.go

@@ -0,0 +1,81 @@
+//go:build darwin
+
+package agent
+
+import (
+	"testing"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestParsePowermetricsData(t *testing.T) {
+	input := `
+Machine model: Mac14,10
+OS version: 25D125
+
+*** Sampled system activity (Sat Feb 14 00:42:06 2026 -0500) (503.05ms elapsed) ***
+
+**** GPU usage ****
+
+GPU HW active frequency: 444 MHz
+GPU HW active residency:   0.97% (444 MHz: .97% 612 MHz:   0% 808 MHz:   0% 968 MHz:   0% 1110 MHz:   0% 1236 MHz:   0% 1338 MHz:   0% 1398 MHz:   0%)
+GPU SW requested state: (P1 : 100% P2 :   0% P3 :   0% P4 :   0% P5 :   0% P6 :   0% P7 :   0% P8 :   0%)
+GPU idle residency:  99.03%
+GPU Power: 4 mW
+`
+	gm := &GPUManager{
+		GpuDataMap: make(map[string]*system.GPUData),
+	}
+	valid := gm.parsePowermetricsData([]byte(input))
+	require.True(t, valid)
+
+	g0, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "Apple GPU", g0.Name)
+	// Usage = 100 - 99.03 = 0.97
+	assert.InDelta(t, 0.97, g0.Usage, 0.01)
+	// 4 mW -> 0.004 W
+	assert.InDelta(t, 0.004, g0.Power, 0.0001)
+	assert.Equal(t, 1.0, g0.Count)
+}
+
+func TestParsePowermetricsDataPartial(t *testing.T) {
+	// Only power line (e.g. older macOS or different sampler output)
+	input := `
+**** GPU usage ****
+GPU Power: 120 mW
+`
+	gm := &GPUManager{
+		GpuDataMap: make(map[string]*system.GPUData),
+	}
+	valid := gm.parsePowermetricsData([]byte(input))
+	require.True(t, valid)
+
+	g0, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "Apple GPU", g0.Name)
+	assert.InDelta(t, 0.12, g0.Power, 0.001)
+	assert.Equal(t, 1.0, g0.Count)
+}
+
+func TestParseMacmonLine(t *testing.T) {
+	input := `{"all_power":0.6468324661254883,"ane_power":0.0,"cpu_power":0.6359732151031494,"ecpu_usage":[2061,0.1726151406764984],"gpu_power":0.010859241709113121,"gpu_ram_power":0.000965250947047025,"gpu_usage":[503,0.013633215799927711],"memory":{"ram_total":17179869184,"ram_usage":12322914304,"swap_total":0,"swap_usage":0},"pcpu_usage":[1248,0.11792058497667313],"ram_power":0.14885640144348145,"sys_power":10.4955415725708,"temp":{"cpu_temp_avg":23.041261672973633,"gpu_temp_avg":29.44516944885254},"timestamp":"2026-02-17T19:34:27.942556+00:00"}`
+
+	gm := &GPUManager{
+		GpuDataMap: make(map[string]*system.GPUData),
+	}
+	valid := gm.parseMacmonLine([]byte(input))
+	require.True(t, valid)
+
+	g0, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "Apple GPU", g0.Name)
+	// macmon reports usage fraction 0..1; expect percent conversion.
+	assert.InDelta(t, 1.3633, g0.Usage, 0.05)
+	// power includes gpu_power + gpu_ram_power
+	assert.InDelta(t, 0.011824, g0.Power, 0.0005)
+	assert.InDelta(t, 29.445, g0.Temperature, 0.01)
+	assert.Equal(t, 1.0, g0.Count)
+}

agent/gpu_nvml_unsupported.go

@@ -13,21 +13,3 @@ func (c *nvmlCollector) init() error {
 }
 
 func (c *nvmlCollector) start() {}
-
-func (c *nvmlCollector) collect() {}
-
-func openLibrary(name string) (uintptr, error) {
-	return 0, fmt.Errorf("nvml not supported on this platform")
-}
-
-func getNVMLPath() string {
-	return ""
-}
-
-func hasSymbol(lib uintptr, symbol string) bool {
-	return false
-}
-
-func (c *nvmlCollector) isGPUActive(bdf string) bool {
-	return true
-}

agent/gpu_nvtop.go

@@ -0,0 +1,159 @@
+package agent
+
+import (
+	"encoding/json"
+	"io"
+	"log/slog"
+	"os/exec"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/henrygd/beszel/internal/entities/system"
+)
+
+type nvtopSnapshot struct {
+	DeviceName string  `json:"device_name"`
+	Temp       *string `json:"temp"`
+	PowerDraw  *string `json:"power_draw"`
+	GpuUtil    *string `json:"gpu_util"`
+	MemTotal   *string `json:"mem_total"`
+	MemUsed    *string `json:"mem_used"`
+}
+
+// parseNvtopNumber parses nvtop numeric strings with units (C/W/%).
+func parseNvtopNumber(raw string) float64 {
+	cleaned := strings.TrimSpace(raw)
+	cleaned = strings.TrimSuffix(cleaned, "C")
+	cleaned = strings.TrimSuffix(cleaned, "W")
+	cleaned = strings.TrimSuffix(cleaned, "%")
+	val, _ := strconv.ParseFloat(cleaned, 64)
+	return val
+}
+
+// parseNvtopData parses a single nvtop JSON snapshot payload.
+func (gm *GPUManager) parseNvtopData(output []byte) bool {
+	var snapshots []nvtopSnapshot
+	if err := json.Unmarshal(output, &snapshots); err != nil || len(snapshots) == 0 {
+		return false
+	}
+	return gm.updateNvtopSnapshots(snapshots)
+}
+
+// updateNvtopSnapshots applies one decoded nvtop snapshot batch to GPU accumulators.
+func (gm *GPUManager) updateNvtopSnapshots(snapshots []nvtopSnapshot) bool {
+	gm.Lock()
+	defer gm.Unlock()
+
+	valid := false
+	usedIDs := make(map[string]struct{}, len(snapshots))
+	for i, sample := range snapshots {
+		if sample.DeviceName == "" {
+			continue
+		}
+		indexID := "n" + strconv.Itoa(i)
+		id := indexID
+
+		// nvtop ordering can change, so prefer reusing an existing slot with matching device name.
+		if existingByIndex, ok := gm.GpuDataMap[indexID]; ok && existingByIndex.Name != "" && existingByIndex.Name != sample.DeviceName {
+			for existingID, gpu := range gm.GpuDataMap {
+				if !strings.HasPrefix(existingID, "n") {
+					continue
+				}
+				if _, taken := usedIDs[existingID]; taken {
+					continue
+				}
+				if gpu.Name == sample.DeviceName {
+					id = existingID
+					break
+				}
+			}
+		}
+
+		if _, ok := gm.GpuDataMap[id]; !ok {
+			gm.GpuDataMap[id] = &system.GPUData{Name: sample.DeviceName}
+		}
+		gpu := gm.GpuDataMap[id]
+		gpu.Name = sample.DeviceName
+
+		if sample.Temp != nil {
+			gpu.Temperature = parseNvtopNumber(*sample.Temp)
+		}
+		if sample.MemUsed != nil {
+			gpu.MemoryUsed = bytesToMegabytes(parseNvtopNumber(*sample.MemUsed))
+		}
+		if sample.MemTotal != nil {
+			gpu.MemoryTotal = bytesToMegabytes(parseNvtopNumber(*sample.MemTotal))
+		}
+		if sample.GpuUtil != nil {
+			gpu.Usage += parseNvtopNumber(*sample.GpuUtil)
+		}
+		if sample.PowerDraw != nil {
+			gpu.Power += parseNvtopNumber(*sample.PowerDraw)
+		}
+		gpu.Count++
+		usedIDs[id] = struct{}{}
+		valid = true
+	}
+	return valid
+}
+
+// collectNvtopStats runs nvtop loop mode and continuously decodes JSON snapshots.
+func (gm *GPUManager) collectNvtopStats(interval string) error {
+	cmd := exec.Command(nvtopCmd, "-lP", "-d", interval)
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return err
+	}
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	defer func() {
+		_ = stdout.Close()
+		if cmd.ProcessState == nil || !cmd.ProcessState.Exited() {
+			_ = cmd.Process.Kill()
+		}
+		_ = cmd.Wait()
+	}()
+
+	decoder := json.NewDecoder(stdout)
+	foundValid := false
+	for {
+		var snapshots []nvtopSnapshot
+		if err := decoder.Decode(&snapshots); err != nil {
+			if err == io.EOF {
+				if foundValid {
+					return nil
+				}
+				return errNoValidData
+			}
+			return err
+		}
+		if gm.updateNvtopSnapshots(snapshots) {
+			foundValid = true
+		}
+	}
+}
+
+// startNvtopCollector starts nvtop collection with retry or fallback callback handling.
+func (gm *GPUManager) startNvtopCollector(interval string, onFailure func()) {
+	go func() {
+		failures := 0
+		for {
+			if err := gm.collectNvtopStats(interval); err != nil {
+				if onFailure != nil {
+					slog.Warn("Error collecting GPU data via nvtop", "err", err)
+					onFailure()
+					return
+				}
+				failures++
+				if failures > maxFailureRetries {
+					break
+				}
+				slog.Warn("Error collecting GPU data via nvtop", "err", err)
+				time.Sleep(retryWaitTime)
+				continue
+			}
+		}
+	}()
+}

agent/gpu_test.go

@@ -250,6 +250,100 @@ func TestParseAmdData(t *testing.T) {
 	}
 }
 
+func TestParseNvtopData(t *testing.T) {
+	input, err := os.ReadFile("test-data/nvtop.json")
+	require.NoError(t, err)
+
+	gm := &GPUManager{
+		GpuDataMap: make(map[string]*system.GPUData),
+	}
+	valid := gm.parseNvtopData(input)
+	require.True(t, valid)
+
+	g0, ok := gm.GpuDataMap["n0"]
+	require.True(t, ok)
+	assert.Equal(t, "NVIDIA GeForce RTX 3050 Ti Laptop GPU", g0.Name)
+	assert.Equal(t, 48.0, g0.Temperature)
+	assert.Equal(t, 5.0, g0.Usage)
+	assert.Equal(t, 13.0, g0.Power)
+	assert.Equal(t, bytesToMegabytes(349372416), g0.MemoryUsed)
+	assert.Equal(t, bytesToMegabytes(4294967296), g0.MemoryTotal)
+	assert.Equal(t, 1.0, g0.Count)
+
+	g1, ok := gm.GpuDataMap["n1"]
+	require.True(t, ok)
+	assert.Equal(t, "AMD Radeon 680M", g1.Name)
+	assert.Equal(t, 48.0, g1.Temperature)
+	assert.Equal(t, 12.0, g1.Usage)
+	assert.Equal(t, 9.0, g1.Power)
+	assert.Equal(t, bytesToMegabytes(1213784064), g1.MemoryUsed)
+	assert.Equal(t, bytesToMegabytes(16929173504), g1.MemoryTotal)
+	assert.Equal(t, 1.0, g1.Count)
+}
+
+func TestUpdateNvtopSnapshotsKeepsDeviceAssociationWhenOrderChanges(t *testing.T) {
+	strPtr := func(s string) *string { return &s }
+
+	gm := &GPUManager{
+		GpuDataMap: make(map[string]*system.GPUData),
+	}
+
+	firstBatch := []nvtopSnapshot{
+		{
+			DeviceName: "NVIDIA GeForce RTX 3050 Ti Laptop GPU",
+			GpuUtil:    strPtr("20%"),
+			PowerDraw:  strPtr("10W"),
+		},
+		{
+			DeviceName: "AMD Radeon 680M",
+			GpuUtil:    strPtr("30%"),
+			PowerDraw:  strPtr("20W"),
+		},
+	}
+	secondBatchSwapped := []nvtopSnapshot{
+		{
+			DeviceName: "AMD Radeon 680M",
+			GpuUtil:    strPtr("40%"),
+			PowerDraw:  strPtr("25W"),
+		},
+		{
+			DeviceName: "NVIDIA GeForce RTX 3050 Ti Laptop GPU",
+			GpuUtil:    strPtr("50%"),
+			PowerDraw:  strPtr("15W"),
+		},
+	}
+
+	require.True(t, gm.updateNvtopSnapshots(firstBatch))
+	require.True(t, gm.updateNvtopSnapshots(secondBatchSwapped))
+
+	nvidia := gm.GpuDataMap["n0"]
+	require.NotNil(t, nvidia)
+	assert.Equal(t, "NVIDIA GeForce RTX 3050 Ti Laptop GPU", nvidia.Name)
+	assert.Equal(t, 70.0, nvidia.Usage)
+	assert.Equal(t, 25.0, nvidia.Power)
+	assert.Equal(t, 2.0, nvidia.Count)
+
+	amd := gm.GpuDataMap["n1"]
+	require.NotNil(t, amd)
+	assert.Equal(t, "AMD Radeon 680M", amd.Name)
+	assert.Equal(t, 70.0, amd.Usage)
+	assert.Equal(t, 45.0, amd.Power)
+	assert.Equal(t, 2.0, amd.Count)
+}
+
+func TestParseCollectorPriority(t *testing.T) {
+	got := parseCollectorPriority(" nvml, nvidia-smi, intel_gpu_top, amd_sysfs, nvtop, rocm-smi, bad ")
+	want := []collectorSource{
+		collectorSourceNVML,
+		collectorSourceNvidiaSMI,
+		collectorSourceIntelGpuTop,
+		collectorSourceAmdSysfs,
+		collectorSourceNVTop,
+		collectorSourceRocmSMI,
+	}
+	assert.Equal(t, want, got)
+}
+
 func TestParseJetsonData(t *testing.T) {
 	tests := []struct {
 		name        string
@@ -987,36 +1081,35 @@ func TestCalculateGPUAverage(t *testing.T) {
 	})
 }
 
-func TestDetectGPUs(t *testing.T) {
+func TestGPUCapabilitiesAndLegacyPriority(t *testing.T) {
 	// Save original PATH
 	origPath := os.Getenv("PATH")
 	defer os.Setenv("PATH", origPath)
-
-	// Set up temp dir with the commands
-	tempDir := t.TempDir()
-	os.Setenv("PATH", tempDir)
+	hasAmdSysfs := (&GPUManager{}).hasAmdSysfs()
 
 	tests := []struct {
 		name           string
-		setupCommands  func() error
+		setupCommands  func(string) error
 		wantNvidiaSmi  bool
 		wantRocmSmi    bool
 		wantTegrastats bool
+		wantNvtop      bool
 		wantErr        bool
 	}{
 		{
 			name: "nvidia-smi not available",
-			setupCommands: func() error {
+			setupCommands: func(_ string) error {
 				return nil
 			},
 			wantNvidiaSmi:  false,
 			wantRocmSmi:    false,
 			wantTegrastats: false,
+			wantNvtop:      false,
 			wantErr:        true,
 		},
 		{
 			name: "nvidia-smi available",
-			setupCommands: func() error {
+			setupCommands: func(tempDir string) error {
 				path := filepath.Join(tempDir, "nvidia-smi")
 				script := `#!/bin/sh
 echo "test"`
@@ -1028,29 +1121,14 @@ echo "test"`
 			wantNvidiaSmi:  true,
 			wantTegrastats: false,
 			wantRocmSmi:    false,
+			wantNvtop:      false,
 			wantErr:        false,
 		},
 		{
 			name: "rocm-smi available",
-			setupCommands: func() error {
+			setupCommands: func(tempDir string) error {
 				path := filepath.Join(tempDir, "rocm-smi")
 				script := `#!/bin/sh
-echo "test"`
-				if err := os.WriteFile(path, []byte(script), 0755); err != nil {
-					return err
-				}
-				return nil
-			},
-			wantNvidiaSmi:  true,
-			wantRocmSmi:    true,
-			wantTegrastats: false,
-			wantErr:        false,
-		},
-		{
-			name: "tegrastats available",
-			setupCommands: func() error {
-				path := filepath.Join(tempDir, "tegrastats")
-				script := `#!/bin/sh
 echo "test"`
 				if err := os.WriteFile(path, []byte(script), 0755); err != nil {
 					return err
@@ -1059,12 +1137,47 @@ echo "test"`
 			},
 			wantNvidiaSmi:  false,
 			wantRocmSmi:    true,
+			wantTegrastats: false,
+			wantNvtop:      false,
+			wantErr:        false,
+		},
+		{
+			name: "tegrastats available",
+			setupCommands: func(tempDir string) error {
+				path := filepath.Join(tempDir, "tegrastats")
+				script := `#!/bin/sh
+echo "test"`
+				if err := os.WriteFile(path, []byte(script), 0755); err != nil {
+					return err
+				}
+				return nil
+			},
+			wantNvidiaSmi:  false,
+			wantRocmSmi:    false,
 			wantTegrastats: true,
+			wantNvtop:      false,
+			wantErr:        false,
+		},
+		{
+			name: "nvtop available",
+			setupCommands: func(tempDir string) error {
+				path := filepath.Join(tempDir, "nvtop")
+				script := `#!/bin/sh
+echo "[]"`
+				if err := os.WriteFile(path, []byte(script), 0755); err != nil {
+					return err
+				}
+				return nil
+			},
+			wantNvidiaSmi:  false,
+			wantRocmSmi:    false,
+			wantTegrastats: false,
+			wantNvtop:      true,
 			wantErr:        false,
 		},
 		{
 			name: "no gpu tools available",
-			setupCommands: func() error {
+			setupCommands: func(_ string) error {
 				os.Setenv("PATH", "")
 				return nil
 			},
@@ -1074,29 +1187,53 @@ echo "test"`
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			if err := tt.setupCommands(); err != nil {
+			tempDir := t.TempDir()
+			os.Setenv("PATH", tempDir)
+			if err := tt.setupCommands(tempDir); err != nil {
 				t.Fatal(err)
 			}
 
 			gm := &GPUManager{}
-			err := gm.detectGPUs()
+			caps := gm.discoverGpuCapabilities()
+			var err error
+			if !hasAnyGpuCollector(caps) {
+				err = fmt.Errorf(noGPUFoundMsg)
+			}
+			priorities := gm.resolveLegacyCollectorPriority(caps)
+			hasPriority := func(source collectorSource) bool {
+				for _, s := range priorities {
+					if s == source {
+						return true
+					}
+				}
+				return false
+			}
+			gotNvidiaSmi := hasPriority(collectorSourceNvidiaSMI)
+			gotRocmSmi := hasPriority(collectorSourceRocmSMI)
+			gotTegrastats := caps.hasTegrastats
+			gotNvtop := caps.hasNvtop
 
-			t.Logf("nvidiaSmi: %v, rocmSmi: %v, tegrastats: %v", gm.nvidiaSmi, gm.rocmSmi, gm.tegrastats)
+			t.Logf("nvidiaSmi: %v, rocmSmi: %v, tegrastats: %v", gotNvidiaSmi, gotRocmSmi, gotTegrastats)
 
-			if tt.wantErr {
+			wantErr := tt.wantErr
+			if hasAmdSysfs && (tt.name == "nvidia-smi not available" || tt.name == "no gpu tools available") {
+				wantErr = false
+			}
+			if wantErr {
 				assert.Error(t, err)
 				return
 			}
 
 			assert.NoError(t, err)
-			assert.Equal(t, tt.wantNvidiaSmi, gm.nvidiaSmi)
-			assert.Equal(t, tt.wantRocmSmi, gm.rocmSmi)
-			assert.Equal(t, tt.wantTegrastats, gm.tegrastats)
+			assert.Equal(t, tt.wantNvidiaSmi, gotNvidiaSmi)
+			assert.Equal(t, tt.wantRocmSmi, gotRocmSmi)
+			assert.Equal(t, tt.wantTegrastats, gotTegrastats)
+			assert.Equal(t, tt.wantNvtop, gotNvtop)
 		})
 	}
 }
 
-func TestStartCollector(t *testing.T) {
+func TestCollectorStartHelpers(t *testing.T) {
 	// Save original PATH
 	origPath := os.Getenv("PATH")
 	defer os.Setenv("PATH", origPath)
@@ -1181,6 +1318,27 @@ echo "11-14-2024 22:54:33 RAM 1024/4096MB GR3D_FREQ 80% tj@70C VDD_GPU_SOC 1000m
 				},
 			},
 		},
+		{
+			name:    "nvtop collector",
+			command: "nvtop",
+			setup: func(t *testing.T) error {
+				path := filepath.Join(dir, "nvtop")
+				script := `#!/bin/sh
+echo '[{"device_name":"NVIDIA Test GPU","temp":"52C","power_draw":"31W","gpu_util":"37%","mem_total":"4294967296","mem_used":"536870912","processes":[]}]'`
+				if err := os.WriteFile(path, []byte(script), 0755); err != nil {
+					return err
+				}
+				return nil
+			},
+			validate: func(t *testing.T, gm *GPUManager) {
+				gpu, exists := gm.GpuDataMap["n0"]
+				assert.True(t, exists)
+				if exists {
+					assert.Equal(t, "NVIDIA Test GPU", gpu.Name)
+					assert.Equal(t, 52.0, gpu.Temperature)
+				}
+			},
+		},
 	}
 
 	for _, tt := range tests {
@@ -1193,13 +1351,157 @@ echo "11-14-2024 22:54:33 RAM 1024/4096MB GR3D_FREQ 80% tj@70C VDD_GPU_SOC 1000m
 					GpuDataMap: make(map[string]*system.GPUData),
 				}
 			}
-			tt.gm.startCollector(tt.command)
+			switch tt.command {
+			case nvidiaSmiCmd:
+				tt.gm.startNvidiaSmiCollector("4")
+			case rocmSmiCmd:
+				tt.gm.startRocmSmiCollector(4300 * time.Millisecond)
+			case tegraStatsCmd:
+				tt.gm.startTegraStatsCollector("3700")
+			case nvtopCmd:
+				tt.gm.startNvtopCollector("30", nil)
+			default:
+				t.Fatalf("unknown test command %q", tt.command)
+			}
 			time.Sleep(50 * time.Millisecond) // Give collector time to run
 			tt.validate(t, tt.gm)
 		})
 	}
 }
 
+func TestNewGPUManagerPriorityNvtopFallback(t *testing.T) {
+	origPath := os.Getenv("PATH")
+	defer os.Setenv("PATH", origPath)
+
+	dir := t.TempDir()
+	os.Setenv("PATH", dir)
+	t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "nvtop,nvidia-smi")
+
+	nvtopPath := filepath.Join(dir, "nvtop")
+	nvtopScript := `#!/bin/sh
+echo 'not-json'`
+	require.NoError(t, os.WriteFile(nvtopPath, []byte(nvtopScript), 0755))
+
+	nvidiaPath := filepath.Join(dir, "nvidia-smi")
+	nvidiaScript := `#!/bin/sh
+echo "0, NVIDIA Priority GPU, 45, 512, 2048, 12, 25"`
+	require.NoError(t, os.WriteFile(nvidiaPath, []byte(nvidiaScript), 0755))
+
+	gm, err := NewGPUManager()
+	require.NoError(t, err)
+	require.NotNil(t, gm)
+
+	time.Sleep(150 * time.Millisecond)
+	gpu, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "Priority GPU", gpu.Name)
+	assert.Equal(t, 45.0, gpu.Temperature)
+}
+
+func TestNewGPUManagerPriorityMixedCollectors(t *testing.T) {
+	origPath := os.Getenv("PATH")
+	defer os.Setenv("PATH", origPath)
+
+	dir := t.TempDir()
+	os.Setenv("PATH", dir)
+	t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "intel_gpu_top,rocm-smi")
+
+	intelPath := filepath.Join(dir, "intel_gpu_top")
+	intelScript := `#!/bin/sh
+echo "Freq MHz      IRQ RC6     Power W     IMC MiB/s             RCS             VCS"
+echo " req  act       /s   %   gpu   pkg     rd     wr       %  se  wa       %  se  wa"
+echo "226  223      338  58  2.00  2.69   1820    965   0.00    0   0    0.00   0   0"
+echo "189  187      412  67  1.80  2.45   1950    823   8.50    2   1    15.00   1   0"
+`
+	require.NoError(t, os.WriteFile(intelPath, []byte(intelScript), 0755))
+
+	rocmPath := filepath.Join(dir, "rocm-smi")
+	rocmScript := `#!/bin/sh
+echo '{"card0": {"Temperature (Sensor edge) (C)": "49.0", "Current Socket Graphics Package Power (W)": "28.159", "GPU use (%)": "0", "VRAM Total Memory (B)": "536870912", "VRAM Total Used Memory (B)": "445550592", "Card Series": "Rembrandt [Radeon 680M]", "GUID": "34756"}}'
+`
+	require.NoError(t, os.WriteFile(rocmPath, []byte(rocmScript), 0755))
+
+	gm, err := NewGPUManager()
+	require.NoError(t, err)
+	require.NotNil(t, gm)
+
+	time.Sleep(150 * time.Millisecond)
+	_, intelOk := gm.GpuDataMap["i0"]
+	_, amdOk := gm.GpuDataMap["34756"]
+	assert.True(t, intelOk)
+	assert.True(t, amdOk)
+}
+
+func TestNewGPUManagerPriorityNvmlFallbackToNvidiaSmi(t *testing.T) {
+	origPath := os.Getenv("PATH")
+	defer os.Setenv("PATH", origPath)
+
+	dir := t.TempDir()
+	os.Setenv("PATH", dir)
+	t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "nvml,nvidia-smi")
+
+	nvidiaPath := filepath.Join(dir, "nvidia-smi")
+	nvidiaScript := `#!/bin/sh
+echo "0, NVIDIA Fallback GPU, 41, 256, 1024, 8, 14"`
+	require.NoError(t, os.WriteFile(nvidiaPath, []byte(nvidiaScript), 0755))
+
+	gm, err := NewGPUManager()
+	require.NoError(t, err)
+	require.NotNil(t, gm)
+
+	time.Sleep(150 * time.Millisecond)
+	gpu, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "Fallback GPU", gpu.Name)
+}
+
+func TestNewGPUManagerConfiguredCollectorsMustStart(t *testing.T) {
+	origPath := os.Getenv("PATH")
+	defer os.Setenv("PATH", origPath)
+
+	dir := t.TempDir()
+	os.Setenv("PATH", dir)
+
+	t.Run("configured valid collector unavailable", func(t *testing.T) {
+		t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "nvidia-smi")
+		gm, err := NewGPUManager()
+		require.Nil(t, gm)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "no configured GPU collectors are available")
+	})
+
+	t.Run("configured collector list has only unknown entries", func(t *testing.T) {
+		t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "bad,unknown")
+		gm, err := NewGPUManager()
+		require.Nil(t, gm)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "no configured GPU collectors are available")
+	})
+}
+
+func TestNewGPUManagerJetsonIgnoresCollectorConfig(t *testing.T) {
+	origPath := os.Getenv("PATH")
+	defer os.Setenv("PATH", origPath)
+
+	dir := t.TempDir()
+	os.Setenv("PATH", dir)
+	t.Setenv("BESZEL_AGENT_GPU_COLLECTOR", "nvidia-smi")
+
+	tegraPath := filepath.Join(dir, "tegrastats")
+	tegraScript := `#!/bin/sh
+echo "11-14-2024 22:54:33 RAM 1024/4096MB GR3D_FREQ 80% tj@70C VDD_GPU_SOC 1000mW"`
+	require.NoError(t, os.WriteFile(tegraPath, []byte(tegraScript), 0755))
+
+	gm, err := NewGPUManager()
+	require.NoError(t, err)
+	require.NotNil(t, gm)
+
+	time.Sleep(100 * time.Millisecond)
+	gpu, ok := gm.GpuDataMap["0"]
+	require.True(t, ok)
+	assert.Equal(t, "GPU", gpu.Name)
+}
+
 // TestAccumulationTableDriven tests the accumulation behavior for all three GPU types
 func TestAccumulation(t *testing.T) {
 	type expectedGPUValues struct {

agent/smart.go

@@ -28,7 +28,7 @@ type SmartManager struct {
 	SmartDevices    []*DeviceInfo
 	refreshMutex    sync.Mutex
 	lastScanTime    time.Time
-	binPath         string
+	smartctlPath    string
 	excludedDevices map[string]struct{}
 }
 
@@ -170,27 +170,35 @@ func (sm *SmartManager) ScanDevices(force bool) error {
 		configuredDevices = parsedDevices
 	}
 
-	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-	defer cancel()
-
-	cmd := exec.CommandContext(ctx, sm.binPath, "--scan", "-j")
-	output, err := cmd.Output()
-
 	var (
 		scanErr        error
 		scannedDevices []*DeviceInfo
 		hasValidScan   bool
 	)
 
-	if err != nil {
-		scanErr = err
-	} else {
-		scannedDevices, hasValidScan = sm.parseScan(output)
-		if !hasValidScan {
-			scanErr = errNoValidSmartData
+	if sm.smartctlPath != "" {
+		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		defer cancel()
+
+		cmd := exec.CommandContext(ctx, sm.smartctlPath, "--scan", "-j")
+		output, err := cmd.Output()
+		if err != nil {
+			scanErr = err
+		} else {
+			scannedDevices, hasValidScan = sm.parseScan(output)
+			if !hasValidScan {
+				scanErr = errNoValidSmartData
+			}
 		}
 	}
 
+	// Add eMMC devices (Linux only) by reading sysfs health fields. This does not
+	// require smartctl and does not scan the whole device.
+	if emmcDevices := scanEmmcDevices(); len(emmcDevices) > 0 {
+		scannedDevices = append(scannedDevices, emmcDevices...)
+		hasValidScan = true
+	}
+
 	finalDevices := mergeDeviceLists(currentDevices, scannedDevices, configuredDevices)
 	finalDevices = sm.filterExcludedDevices(finalDevices)
 	sm.updateSmartDevices(finalDevices)
@@ -442,6 +450,18 @@ func (sm *SmartManager) CollectSmart(deviceInfo *DeviceInfo) error {
 		return errNoValidSmartData
 	}
 
+	// eMMC health is not exposed via SMART on Linux, but the kernel provides
+	// wear / EOL indicators via sysfs. Prefer that path when available.
+	if deviceInfo != nil {
+		if ok, err := sm.collectEmmcHealth(deviceInfo); ok {
+			return err
+		}
+	}
+
+	if sm.smartctlPath == "" {
+		return errNoValidSmartData
+	}
+
 	// slog.Info("collecting SMART data", "device", deviceInfo.Name, "type", deviceInfo.Type, "has_existing_data", sm.hasDataForDevice(deviceInfo.Name))
 
 	// Check if we have any existing data for this device
@@ -452,7 +472,7 @@ func (sm *SmartManager) CollectSmart(deviceInfo *DeviceInfo) error {
 
 	// Try with -n standby first if we have existing data
 	args := sm.smartctlArgs(deviceInfo, hasExistingData)
-	cmd := exec.CommandContext(ctx, sm.binPath, args...)
+	cmd := exec.CommandContext(ctx, sm.smartctlPath, args...)
 	output, err := cmd.CombinedOutput()
 
 	// Check if device is in standby (exit status 2)
@@ -465,7 +485,7 @@ func (sm *SmartManager) CollectSmart(deviceInfo *DeviceInfo) error {
 		ctx2, cancel2 := context.WithTimeout(context.Background(), 15*time.Second)
 		defer cancel2()
 		args = sm.smartctlArgs(deviceInfo, false)
-		cmd = exec.CommandContext(ctx2, sm.binPath, args...)
+		cmd = exec.CommandContext(ctx2, sm.smartctlPath, args...)
 		output, err = cmd.CombinedOutput()
 	}
 
@@ -482,7 +502,7 @@ func (sm *SmartManager) CollectSmart(deviceInfo *DeviceInfo) error {
 			ctx3, cancel3 := context.WithTimeout(context.Background(), 15*time.Second)
 			defer cancel3()
 			args = sm.smartctlArgs(deviceInfo, false)
-			cmd = exec.CommandContext(ctx3, sm.binPath, args...)
+			cmd = exec.CommandContext(ctx3, sm.smartctlPath, args...)
 			output, err = cmd.CombinedOutput()
 			hasValidData = sm.parseSmartOutput(deviceInfo, output)
 
@@ -1123,10 +1143,15 @@ func NewSmartManager() (*SmartManager, error) {
 	}
 	sm.refreshExcludedDevices()
 	path, err := sm.detectSmartctl()
+	slog.Debug("smartctl", "path", path, "err", err)
 	if err != nil {
+		// Keep the previous fail-fast behavior unless this Linux host exposes
+		// eMMC health via sysfs, in which case smartctl is optional.
+		if runtime.GOOS == "linux" && len(scanEmmcDevices()) > 0 {
+			return sm, nil
+		}
 		return nil, err
 	}
-	slog.Debug("smartctl", "path", path)
-	sm.binPath = path
+	sm.smartctlPath = path
 	return sm, nil
 }

agent/test-data/amdgpu.ids

@@ -0,0 +1,700 @@
+# List of AMDGPU IDs
+#
+# Syntax:
+# device_id,	revision_id,	product_name        <-- single tab after comma
+
+1.0.0
+1114,	C2,	AMD Radeon 860M Graphics
+1114,	C3,	AMD Radeon 840M Graphics
+1114,	D2,	AMD Radeon 860M Graphics
+1114,	D3,	AMD Radeon 840M Graphics
+1309,	00,	AMD Radeon R7 Graphics
+130A,	00,	AMD Radeon R6 Graphics
+130B,	00,	AMD Radeon R4 Graphics
+130C,	00,	AMD Radeon R7 Graphics
+130D,	00,	AMD Radeon R6 Graphics
+130E,	00,	AMD Radeon R5 Graphics
+130F,	00,	AMD Radeon R7 Graphics
+130F,	D4,	AMD Radeon R7 Graphics
+130F,	D5,	AMD Radeon R7 Graphics
+130F,	D6,	AMD Radeon R7 Graphics
+130F,	D7,	AMD Radeon R7 Graphics
+1313,	00,	AMD Radeon R7 Graphics
+1313,	D4,	AMD Radeon R7 Graphics
+1313,	D5,	AMD Radeon R7 Graphics
+1313,	D6,	AMD Radeon R7 Graphics
+1315,	00,	AMD Radeon R5 Graphics
+1315,	D4,	AMD Radeon R5 Graphics
+1315,	D5,	AMD Radeon R5 Graphics
+1315,	D6,	AMD Radeon R5 Graphics
+1315,	D7,	AMD Radeon R5 Graphics
+1316,	00,	AMD Radeon R5 Graphics
+1318,	00,	AMD Radeon R5 Graphics
+131B,	00,	AMD Radeon R4 Graphics
+131C,	00,	AMD Radeon R7 Graphics
+131D,	00,	AMD Radeon R6 Graphics
+1435,	AE,	AMD Custom GPU 0932
+1506,	C1,	AMD Radeon 610M
+1506,	C2,	AMD Radeon 610M
+1506,	C3,	AMD Radeon 610M
+1506,	C4,	AMD Radeon 610M
+150E,	C1,	AMD Radeon 890M Graphics
+150E,	C4,	AMD Radeon 890M Graphics
+150E,	C5,	AMD Radeon 890M Graphics
+150E,	C6,	AMD Radeon 890M Graphics
+150E,	D1,	AMD Radeon 890M Graphics
+150E,	D2,	AMD Radeon 890M Graphics
+150E,	D3,	AMD Radeon 890M Graphics
+1586,	C1,	Radeon 8060S Graphics
+1586,	C2,	Radeon 8050S Graphics
+1586,	C4,	Radeon 8050S Graphics
+1586,	D1,	Radeon 8060S Graphics
+1586,	D2,	Radeon 8050S Graphics
+1586,	D4,	Radeon 8050S Graphics
+1586,	D5,	Radeon 8040S Graphics
+15BF,	00,	AMD Radeon 780M Graphics
+15BF,	01,	AMD Radeon 760M Graphics
+15BF,	02,	AMD Radeon 780M Graphics
+15BF,	03,	AMD Radeon 760M Graphics
+15BF,	C1,	AMD Radeon 780M Graphics
+15BF,	C2,	AMD Radeon 780M Graphics
+15BF,	C3,	AMD Radeon 760M Graphics
+15BF,	C4,	AMD Radeon 780M Graphics
+15BF,	C5,	AMD Radeon 740M Graphics
+15BF,	C6,	AMD Radeon 780M Graphics
+15BF,	C7,	AMD Radeon 780M Graphics
+15BF,	C8,	AMD Radeon 760M Graphics
+15BF,	C9,	AMD Radeon 780M Graphics
+15BF,	CA,	AMD Radeon 740M Graphics
+15BF,	CB,	AMD Radeon 760M Graphics
+15BF,	CC,	AMD Radeon 740M Graphics
+15BF,	CD,	AMD Radeon 760M Graphics
+15BF,	CF,	AMD Radeon 780M Graphics
+15BF,	D0,	AMD Radeon 780M Graphics
+15BF,	D1,	AMD Radeon 780M Graphics
+15BF,	D2,	AMD Radeon 780M Graphics
+15BF,	D3,	AMD Radeon 780M Graphics
+15BF,	D4,	AMD Radeon 780M Graphics
+15BF,	D5,	AMD Radeon 760M Graphics
+15BF,	D6,	AMD Radeon 760M Graphics
+15BF,	D7,	AMD Radeon 780M Graphics
+15BF,	D8,	AMD Radeon 740M Graphics
+15BF,	D9,	AMD Radeon 780M Graphics
+15BF,	DA,	AMD Radeon 780M Graphics
+15BF,	DB,	AMD Radeon 760M Graphics
+15BF,	DC,	AMD Radeon 760M Graphics
+15BF,	DD,	AMD Radeon 780M Graphics
+15BF,	DE,	AMD Radeon 740M Graphics
+15BF,	DF,	AMD Radeon 760M Graphics
+15BF,	F0,	AMD Radeon 760M Graphics
+15C8,	C1,	AMD Radeon 740M Graphics
+15C8,	C2,	AMD Radeon 740M Graphics
+15C8,	C3,	AMD Radeon 740M Graphics
+15C8,	C4,	AMD Radeon 740M Graphics
+15C8,	D1,	AMD Radeon 740M Graphics
+15C8,	D2,	AMD Radeon 740M Graphics
+15C8,	D3,	AMD Radeon 740M Graphics
+15C8,	D4,	AMD Radeon 740M Graphics
+15D8,	00,	AMD Radeon RX Vega 8 Graphics WS
+15D8,	91,	AMD Radeon Vega 3 Graphics
+15D8,	91,	AMD Ryzen Embedded R1606G with Radeon Vega Gfx
+15D8,	92,	AMD Radeon Vega 3 Graphics
+15D8,	92,	AMD Ryzen Embedded R1505G with Radeon Vega Gfx
+15D8,	93,	AMD Radeon Vega 1 Graphics
+15D8,	A1,	AMD Radeon Vega 10 Graphics
+15D8,	A2,	AMD Radeon Vega 8 Graphics
+15D8,	A3,	AMD Radeon Vega 6 Graphics
+15D8,	A4,	AMD Radeon Vega 3 Graphics
+15D8,	B1,	AMD Radeon Vega 10 Graphics
+15D8,	B2,	AMD Radeon Vega 8 Graphics
+15D8,	B3,	AMD Radeon Vega 6 Graphics
+15D8,	B4,	AMD Radeon Vega 3 Graphics
+15D8,	C1,	AMD Radeon Vega 10 Graphics
+15D8,	C2,	AMD Radeon Vega 8 Graphics
+15D8,	C3,	AMD Radeon Vega 6 Graphics
+15D8,	C4,	AMD Radeon Vega 3 Graphics
+15D8,	C5,	AMD Radeon Vega 3 Graphics
+15D8,	C8,	AMD Radeon Vega 11 Graphics
+15D8,	C9,	AMD Radeon Vega 8 Graphics
+15D8,	CA,	AMD Radeon Vega 11 Graphics
+15D8,	CB,	AMD Radeon Vega 8 Graphics
+15D8,	CC,	AMD Radeon Vega 3 Graphics
+15D8,	CE,	AMD Radeon Vega 3 Graphics
+15D8,	CF,	AMD Ryzen Embedded R1305G with Radeon Vega Gfx
+15D8,	D1,	AMD Radeon Vega 10 Graphics
+15D8,	D2,	AMD Radeon Vega 8 Graphics
+15D8,	D3,	AMD Radeon Vega 6 Graphics
+15D8,	D4,	AMD Radeon Vega 3 Graphics
+15D8,	D8,	AMD Radeon Vega 11 Graphics
+15D8,	D9,	AMD Radeon Vega 8 Graphics
+15D8,	DA,	AMD Radeon Vega 11 Graphics
+15D8,	DB,	AMD Radeon Vega 3 Graphics
+15D8,	DB,	AMD Radeon Vega 8 Graphics
+15D8,	DC,	AMD Radeon Vega 3 Graphics
+15D8,	DD,	AMD Radeon Vega 3 Graphics
+15D8,	DE,	AMD Radeon Vega 3 Graphics
+15D8,	DF,	AMD Radeon Vega 3 Graphics
+15D8,	E3,	AMD Radeon Vega 3 Graphics
+15D8,	E4,	AMD Ryzen Embedded R1102G with Radeon Vega Gfx
+15DD,	81,	AMD Ryzen Embedded V1807B with Radeon Vega Gfx
+15DD,	82,	AMD Ryzen Embedded V1756B with Radeon Vega Gfx
+15DD,	83,	AMD Ryzen Embedded V1605B with Radeon Vega Gfx
+15DD,	84,	AMD Radeon Vega 6 Graphics
+15DD,	85,	AMD Ryzen Embedded V1202B with Radeon Vega Gfx
+15DD,	86,	AMD Radeon Vega 11 Graphics
+15DD,	88,	AMD Radeon Vega 8 Graphics
+15DD,	C1,	AMD Radeon Vega 11 Graphics
+15DD,	C2,	AMD Radeon Vega 8 Graphics
+15DD,	C3,	AMD Radeon Vega 3 / 10 Graphics
+15DD,	C4,	AMD Radeon Vega 8 Graphics
+15DD,	C5,	AMD Radeon Vega 3 Graphics
+15DD,	C6,	AMD Radeon Vega 11 Graphics
+15DD,	C8,	AMD Radeon Vega 8 Graphics
+15DD,	C9,	AMD Radeon Vega 11 Graphics
+15DD,	CA,	AMD Radeon Vega 8 Graphics
+15DD,	CB,	AMD Radeon Vega 3 Graphics
+15DD,	CC,	AMD Radeon Vega 6 Graphics
+15DD,	CE,	AMD Radeon Vega 3 Graphics
+15DD,	CF,	AMD Radeon Vega 3 Graphics
+15DD,	D0,	AMD Radeon Vega 10 Graphics
+15DD,	D1,	AMD Radeon Vega 8 Graphics
+15DD,	D3,	AMD Radeon Vega 11 Graphics
+15DD,	D5,	AMD Radeon Vega 8 Graphics
+15DD,	D6,	AMD Radeon Vega 11 Graphics
+15DD,	D7,	AMD Radeon Vega 8 Graphics
+15DD,	D8,	AMD Radeon Vega 3 Graphics
+15DD,	D9,	AMD Radeon Vega 6 Graphics
+15DD,	E1,	AMD Radeon Vega 3 Graphics
+15DD,	E2,	AMD Radeon Vega 3 Graphics
+163F,	AE,	AMD Custom GPU 0405
+163F,	E1,	AMD Custom GPU 0405
+164E,	D8,	AMD Radeon 610M
+164E,	D9,	AMD Radeon 610M
+164E,	DA,	AMD Radeon 610M
+164E,	DB,	AMD Radeon 610M
+164E,	DC,	AMD Radeon 610M
+1681,	06,	AMD Radeon 680M
+1681,	07,	AMD Radeon 660M
+1681,	0A,	AMD Radeon 680M
+1681,	0B,	AMD Radeon 660M
+1681,	C7,	AMD Radeon 680M
+1681,	C8,	AMD Radeon 680M
+1681,	C9,	AMD Radeon 660M
+1900,	01,	AMD Radeon 780M Graphics
+1900,	02,	AMD Radeon 760M Graphics
+1900,	03,	AMD Radeon 780M Graphics
+1900,	04,	AMD Radeon 760M Graphics
+1900,	05,	AMD Radeon 780M Graphics
+1900,	06,	AMD Radeon 780M Graphics
+1900,	07,	AMD Radeon 760M Graphics
+1900,	B0,	AMD Radeon 780M Graphics
+1900,	B1,	AMD Radeon 780M Graphics
+1900,	B2,	AMD Radeon 780M Graphics
+1900,	B3,	AMD Radeon 780M Graphics
+1900,	B4,	AMD Radeon 780M Graphics
+1900,	B5,	AMD Radeon 780M Graphics
+1900,	B6,	AMD Radeon 780M Graphics
+1900,	B7,	AMD Radeon 760M Graphics
+1900,	B8,	AMD Radeon 760M Graphics
+1900,	B9,	AMD Radeon 780M Graphics
+1900,	BA,	AMD Radeon 780M Graphics
+1900,	BB,	AMD Radeon 780M Graphics
+1900,	C0,	AMD Radeon 780M Graphics
+1900,	C1,	AMD Radeon 760M Graphics
+1900,	C2,	AMD Radeon 780M Graphics
+1900,	C3,	AMD Radeon 760M Graphics
+1900,	C4,	AMD Radeon 780M Graphics
+1900,	C5,	AMD Radeon 780M Graphics
+1900,	C6,	AMD Radeon 760M Graphics
+1900,	C7,	AMD Radeon 780M Graphics
+1900,	C8,	AMD Radeon 760M Graphics
+1900,	C9,	AMD Radeon 780M Graphics
+1900,	CA,	AMD Radeon 760M Graphics
+1900,	CB,	AMD Radeon 780M Graphics
+1900,	CC,	AMD Radeon 780M Graphics
+1900,	CD,	AMD Radeon 760M Graphics
+1900,	CE,	AMD Radeon 780M Graphics
+1900,	CF,	AMD Radeon 760M Graphics
+1900,	D0,	AMD Radeon 780M Graphics
+1900,	D1,	AMD Radeon 760M Graphics
+1900,	D2,	AMD Radeon 780M Graphics
+1900,	D3,	AMD Radeon 760M Graphics
+1900,	D4,	AMD Radeon 780M Graphics
+1900,	D5,	AMD Radeon 780M Graphics
+1900,	D6,	AMD Radeon 760M Graphics
+1900,	D7,	AMD Radeon 780M Graphics
+1900,	D8,	AMD Radeon 760M Graphics
+1900,	D9,	AMD Radeon 780M Graphics
+1900,	DA,	AMD Radeon 760M Graphics
+1900,	DB,	AMD Radeon 780M Graphics
+1900,	DC,	AMD Radeon 780M Graphics
+1900,	DD,	AMD Radeon 760M Graphics
+1900,	DE,	AMD Radeon 780M Graphics
+1900,	DF,	AMD Radeon 760M Graphics
+1900,	F0,	AMD Radeon 780M Graphics
+1900,	F1,	AMD Radeon 780M Graphics
+1900,	F2,	AMD Radeon 780M Graphics
+1901,	C1,	AMD Radeon 740M Graphics
+1901,	C2,	AMD Radeon 740M Graphics
+1901,	C3,	AMD Radeon 740M Graphics
+1901,	C6,	AMD Radeon 740M Graphics
+1901,	C7,	AMD Radeon 740M Graphics
+1901,	C8,	AMD Radeon 740M Graphics
+1901,	C9,	AMD Radeon 740M Graphics
+1901,	CA,	AMD Radeon 740M Graphics
+1901,	D1,	AMD Radeon 740M Graphics
+1901,	D2,	AMD Radeon 740M Graphics
+1901,	D3,	AMD Radeon 740M Graphics
+1901,	D4,	AMD Radeon 740M Graphics
+1901,	D5,	AMD Radeon 740M Graphics
+1901,	D6,	AMD Radeon 740M Graphics
+1901,	D7,	AMD Radeon 740M Graphics
+1901,	D8,	AMD Radeon 740M Graphics
+6600,	00,	AMD Radeon HD 8600 / 8700M
+6600,	81,	AMD Radeon R7 M370
+6601,	00,	AMD Radeon HD 8500M / 8700M
+6604,	00,	AMD Radeon R7 M265 Series
+6604,	81,	AMD Radeon R7 M350
+6605,	00,	AMD Radeon R7 M260 Series
+6605,	81,	AMD Radeon R7 M340
+6606,	00,	AMD Radeon HD 8790M
+6607,	00,	AMD Radeon R5 M240
+6608,	00,	AMD FirePro W2100
+6610,	00,	AMD Radeon R7 200 Series
+6610,	81,	AMD Radeon R7 350
+6610,	83,	AMD Radeon R5 340
+6610,	87,	AMD Radeon R7 200 Series
+6611,	00,	AMD Radeon R7 200 Series
+6611,	87,	AMD Radeon R7 200 Series
+6613,	00,	AMD Radeon R7 200 Series
+6617,	00,	AMD Radeon R7 240 Series
+6617,	87,	AMD Radeon R7 200 Series
+6617,	C7,	AMD Radeon R7 240 Series
+6640,	00,	AMD Radeon HD 8950
+6640,	80,	AMD Radeon R9 M380
+6646,	00,	AMD Radeon R9 M280X
+6646,	80,	AMD Radeon R9 M385
+6646,	80,	AMD Radeon R9 M470X
+6647,	00,	AMD Radeon R9 M200X Series
+6647,	80,	AMD Radeon R9 M380
+6649,	00,	AMD FirePro W5100
+6658,	00,	AMD Radeon R7 200 Series
+665C,	00,	AMD Radeon HD 7700 Series
+665D,	00,	AMD Radeon R7 200 Series
+665F,	81,	AMD Radeon R7 360 Series
+6660,	00,	AMD Radeon HD 8600M Series
+6660,	81,	AMD Radeon R5 M335
+6660,	83,	AMD Radeon R5 M330
+6663,	00,	AMD Radeon HD 8500M Series
+6663,	83,	AMD Radeon R5 M320
+6664,	00,	AMD Radeon R5 M200 Series
+6665,	00,	AMD Radeon R5 M230 Series
+6665,	83,	AMD Radeon R5 M320
+6665,	C3,	AMD Radeon R5 M435
+6666,	00,	AMD Radeon R5 M200 Series
+6667,	00,	AMD Radeon R5 M200 Series
+666F,	00,	AMD Radeon HD 8500M
+66A1,	02,	AMD Instinct MI60 / MI50
+66A1,	06,	AMD Radeon Pro VII
+66AF,	C1,	AMD Radeon VII
+6780,	00,	AMD FirePro W9000
+6784,	00,	ATI FirePro V (FireGL V) Graphics Adapter
+6788,	00,	ATI FirePro V (FireGL V) Graphics Adapter
+678A,	00,	AMD FirePro W8000
+6798,	00,	AMD Radeon R9 200 / HD 7900 Series
+6799,	00,	AMD Radeon HD 7900 Series
+679A,	00,	AMD Radeon HD 7900 Series
+679B,	00,	AMD Radeon HD 7900 Series
+679E,	00,	AMD Radeon HD 7800 Series
+67A0,	00,	AMD Radeon FirePro W9100
+67A1,	00,	AMD Radeon FirePro W8100
+67B0,	00,	AMD Radeon R9 200 Series
+67B0,	80,	AMD Radeon R9 390 Series
+67B1,	00,	AMD Radeon R9 200 Series
+67B1,	80,	AMD Radeon R9 390 Series
+67B9,	00,	AMD Radeon R9 200 Series
+67C0,	00,	AMD Radeon Pro WX 7100 Graphics
+67C0,	80,	AMD Radeon E9550
+67C2,	01,	AMD Radeon Pro V7350x2
+67C2,	02,	AMD Radeon Pro V7300X
+67C4,	00,	AMD Radeon Pro WX 7100 Graphics
+67C4,	80,	AMD Radeon E9560 / E9565 Graphics
+67C7,	00,	AMD Radeon Pro WX 5100 Graphics
+67C7,	80,	AMD Radeon E9390 Graphics
+67D0,	01,	AMD Radeon Pro V7350x2
+67D0,	02,	AMD Radeon Pro V7300X
+67DF,	C0,	AMD Radeon Pro 580X
+67DF,	C1,	AMD Radeon RX 580 Series
+67DF,	C2,	AMD Radeon RX 570 Series
+67DF,	C3,	AMD Radeon RX 580 Series
+67DF,	C4,	AMD Radeon RX 480 Graphics
+67DF,	C5,	AMD Radeon RX 470 Graphics
+67DF,	C6,	AMD Radeon RX 570 Series
+67DF,	C7,	AMD Radeon RX 480 Graphics
+67DF,	CF,	AMD Radeon RX 470 Graphics
+67DF,	D7,	AMD Radeon RX 470 Graphics
+67DF,	E0,	AMD Radeon RX 470 Series
+67DF,	E1,	AMD Radeon RX 590 Series
+67DF,	E3,	AMD Radeon RX Series
+67DF,	E7,	AMD Radeon RX 580 Series
+67DF,	EB,	AMD Radeon Pro 580X
+67DF,	EF,	AMD Radeon RX 570 Series
+67DF,	F7,	AMD Radeon RX P30PH
+67DF,	FF,	AMD Radeon RX 470 Series
+67E0,	00,	AMD Radeon Pro WX Series
+67E3,	00,	AMD Radeon Pro WX 4100
+67E8,	00,	AMD Radeon Pro WX Series
+67E8,	01,	AMD Radeon Pro WX Series
+67E8,	80,	AMD Radeon E9260 Graphics
+67EB,	00,	AMD Radeon Pro V5300X
+67EF,	C0,	AMD Radeon RX Graphics
+67EF,	C1,	AMD Radeon RX 460 Graphics
+67EF,	C2,	AMD Radeon Pro Series
+67EF,	C3,	AMD Radeon RX Series
+67EF,	C5,	AMD Radeon RX 460 Graphics
+67EF,	C7,	AMD Radeon RX Graphics
+67EF,	CF,	AMD Radeon RX 460 Graphics
+67EF,	E0,	AMD Radeon RX 560 Series
+67EF,	E1,	AMD Radeon RX Series
+67EF,	E2,	AMD Radeon RX 560X
+67EF,	E3,	AMD Radeon RX Series
+67EF,	E5,	AMD Radeon RX 560 Series
+67EF,	E7,	AMD Radeon RX 560 Series
+67EF,	EF,	AMD Radeon 550 Series
+67EF,	FF,	AMD Radeon RX 460 Graphics
+67FF,	C0,	AMD Radeon Pro 465
+67FF,	C1,	AMD Radeon RX 560 Series
+67FF,	CF,	AMD Radeon RX 560 Series
+67FF,	EF,	AMD Radeon RX 560 Series
+67FF,	FF,	AMD Radeon RX 550 Series
+6800,	00,	AMD Radeon HD 7970M
+6801,	00,	AMD Radeon HD 8970M
+6806,	00,	AMD Radeon R9 M290X
+6808,	00,	AMD FirePro W7000
+6808,	00,	ATI FirePro V (FireGL V) Graphics Adapter
+6809,	00,	ATI FirePro W5000
+6810,	00,	AMD Radeon R9 200 Series
+6810,	81,	AMD Radeon R9 370 Series
+6811,	00,	AMD Radeon R9 200 Series
+6811,	81,	AMD Radeon R7 370 Series
+6818,	00,	AMD Radeon HD 7800 Series
+6819,	00,	AMD Radeon HD 7800 Series
+6820,	00,	AMD Radeon R9 M275X
+6820,	81,	AMD Radeon R9 M375
+6820,	83,	AMD Radeon R9 M375X
+6821,	00,	AMD Radeon R9 M200X Series
+6821,	83,	AMD Radeon R9 M370X
+6821,	87,	AMD Radeon R7 M380
+6822,	00,	AMD Radeon E8860
+6823,	00,	AMD Radeon R9 M200X Series
+6825,	00,	AMD Radeon HD 7800M Series
+6826,	00,	AMD Radeon HD 7700M Series
+6827,	00,	AMD Radeon HD 7800M Series
+6828,	00,	AMD FirePro W600
+682B,	00,	AMD Radeon HD 8800M Series
+682B,	87,	AMD Radeon R9 M360
+682C,	00,	AMD FirePro W4100
+682D,	00,	AMD Radeon HD 7700M Series
+682F,	00,	AMD Radeon HD 7700M Series
+6830,	00,	AMD Radeon 7800M Series
+6831,	00,	AMD Radeon 7700M Series
+6835,	00,	AMD Radeon R7 Series / HD 9000 Series
+6837,	00,	AMD Radeon HD 7700 Series
+683D,	00,	AMD Radeon HD 7700 Series
+683F,	00,	AMD Radeon HD 7700 Series
+684C,	00,	ATI FirePro V (FireGL V) Graphics Adapter
+6860,	00,	AMD Radeon Instinct MI25
+6860,	01,	AMD Radeon Instinct MI25
+6860,	02,	AMD Radeon Instinct MI25
+6860,	03,	AMD Radeon Pro V340
+6860,	04,	AMD Radeon Instinct MI25x2
+6860,	07,	AMD Radeon Pro V320
+6861,	00,	AMD Radeon Pro WX 9100
+6862,	00,	AMD Radeon Pro SSG
+6863,	00,	AMD Radeon Vega Frontier Edition
+6864,	03,	AMD Radeon Pro V340
+6864,	04,	AMD Radeon Instinct MI25x2
+6864,	05,	AMD Radeon Pro V340
+6868,	00,	AMD Radeon Pro WX 8200
+686C,	00,	AMD Radeon Instinct MI25 MxGPU
+686C,	01,	AMD Radeon Instinct MI25 MxGPU
+686C,	02,	AMD Radeon Instinct MI25 MxGPU
+686C,	03,	AMD Radeon Pro V340 MxGPU
+686C,	04,	AMD Radeon Instinct MI25x2 MxGPU
+686C,	05,	AMD Radeon Pro V340L MxGPU
+686C,	06,	AMD Radeon Instinct MI25 MxGPU
+687F,	01,	AMD Radeon RX Vega
+687F,	C0,	AMD Radeon RX Vega
+687F,	C1,	AMD Radeon RX Vega
+687F,	C3,	AMD Radeon RX Vega
+687F,	C7,	AMD Radeon RX Vega
+6900,	00,	AMD Radeon R7 M260
+6900,	81,	AMD Radeon R7 M360
+6900,	83,	AMD Radeon R7 M340
+6900,	C1,	AMD Radeon R5 M465 Series
+6900,	C3,	AMD Radeon R5 M445 Series
+6900,	D1,	AMD Radeon 530 Series
+6900,	D3,	AMD Radeon 530 Series
+6901,	00,	AMD Radeon R5 M255
+6902,	00,	AMD Radeon Series
+6907,	00,	AMD Radeon R5 M255
+6907,	87,	AMD Radeon R5 M315
+6920,	00,	AMD Radeon R9 M395X
+6920,	01,	AMD Radeon R9 M390X
+6921,	00,	AMD Radeon R9 M390X
+6929,	00,	AMD FirePro S7150
+6929,	01,	AMD FirePro S7100X
+692B,	00,	AMD FirePro W7100
+6938,	00,	AMD Radeon R9 200 Series
+6938,	F0,	AMD Radeon R9 200 Series
+6938,	F1,	AMD Radeon R9 380 Series
+6939,	00,	AMD Radeon R9 200 Series
+6939,	F0,	AMD Radeon R9 200 Series
+6939,	F1,	AMD Radeon R9 380 Series
+694C,	C0,	AMD Radeon RX Vega M GH Graphics
+694E,	C0,	AMD Radeon RX Vega M GL Graphics
+6980,	00,	AMD Radeon Pro WX 3100
+6981,	00,	AMD Radeon Pro WX 3200 Series
+6981,	01,	AMD Radeon Pro WX 3200 Series
+6981,	10,	AMD Radeon Pro WX 3200 Series
+6985,	00,	AMD Radeon Pro WX 3100
+6986,	00,	AMD Radeon Pro WX 2100
+6987,	80,	AMD Embedded Radeon E9171
+6987,	C0,	AMD Radeon 550X Series
+6987,	C1,	AMD Radeon RX 640
+6987,	C3,	AMD Radeon 540X Series
+6987,	C7,	AMD Radeon 540
+6995,	00,	AMD Radeon Pro WX 2100
+6997,	00,	AMD Radeon Pro WX 2100
+699F,	81,	AMD Embedded Radeon E9170 Series
+699F,	C0,	AMD Radeon 500 Series
+699F,	C1,	AMD Radeon 540 Series
+699F,	C3,	AMD Radeon 500 Series
+699F,	C7,	AMD Radeon RX 550 / 550 Series
+699F,	C9,	AMD Radeon 540
+6FDF,	E7,	AMD Radeon RX 590 GME
+6FDF,	EF,	AMD Radeon RX 580 2048SP
+7300,	C1,	AMD FirePro S9300 x2
+7300,	C8,	AMD Radeon R9 Fury Series
+7300,	C9,	AMD Radeon Pro Duo
+7300,	CA,	AMD Radeon R9 Fury Series
+7300,	CB,	AMD Radeon R9 Fury Series
+7312,	00,	AMD Radeon Pro W5700
+731E,	C6,	AMD Radeon RX 5700XTB
+731E,	C7,	AMD Radeon RX 5700B
+731F,	C0,	AMD Radeon RX 5700 XT 50th Anniversary
+731F,	C1,	AMD Radeon RX 5700 XT
+731F,	C2,	AMD Radeon RX 5600M
+731F,	C3,	AMD Radeon RX 5700M
+731F,	C4,	AMD Radeon RX 5700
+731F,	C5,	AMD Radeon RX 5700 XT
+731F,	CA,	AMD Radeon RX 5600 XT
+731F,	CB,	AMD Radeon RX 5600 OEM
+7340,	C1,	AMD Radeon RX 5500M
+7340,	C3,	AMD Radeon RX 5300M
+7340,	C5,	AMD Radeon RX 5500 XT
+7340,	C7,	AMD Radeon RX 5500
+7340,	C9,	AMD Radeon RX 5500XTB
+7340,	CF,	AMD Radeon RX 5300
+7341,	00,	AMD Radeon Pro W5500
+7347,	00,	AMD Radeon Pro W5500M
+7360,	41,	AMD Radeon Pro 5600M
+7360,	C3,	AMD Radeon Pro V520
+7362,	C1,	AMD Radeon Pro V540
+7362,	C3,	AMD Radeon Pro V520
+738C,	01,	AMD Instinct MI100
+73A1,	00,	AMD Radeon Pro V620
+73A3,	00,	AMD Radeon Pro W6800
+73A5,	C0,	AMD Radeon RX 6950 XT
+73AE,	00,	AMD Radeon Pro V620 MxGPU
+73AF,	C0,	AMD Radeon RX 6900 XT
+73BF,	C0,	AMD Radeon RX 6900 XT
+73BF,	C1,	AMD Radeon RX 6800 XT
+73BF,	C3,	AMD Radeon RX 6800
+73DF,	C0,	AMD Radeon RX 6750 XT
+73DF,	C1,	AMD Radeon RX 6700 XT
+73DF,	C2,	AMD Radeon RX 6800M
+73DF,	C3,	AMD Radeon RX 6800M
+73DF,	C5,	AMD Radeon RX 6700 XT
+73DF,	CF,	AMD Radeon RX 6700M
+73DF,	D5,	AMD Radeon RX 6750 GRE 12GB
+73DF,	D7,	AMD TDC-235
+73DF,	DF,	AMD Radeon RX 6700
+73DF,	E5,	AMD Radeon RX 6750 GRE 12GB
+73DF,	FF,	AMD Radeon RX 6700
+73E0,	00,	AMD Radeon RX 6600M
+73E1,	00,	AMD Radeon Pro W6600M
+73E3,	00,	AMD Radeon Pro W6600
+73EF,	C0,	AMD Radeon RX 6800S
+73EF,	C1,	AMD Radeon RX 6650 XT
+73EF,	C2,	AMD Radeon RX 6700S
+73EF,	C3,	AMD Radeon RX 6650M
+73EF,	C4,	AMD Radeon RX 6650M XT
+73FF,	C1,	AMD Radeon RX 6600 XT
+73FF,	C3,	AMD Radeon RX 6600M
+73FF,	C7,	AMD Radeon RX 6600
+73FF,	CB,	AMD Radeon RX 6600S
+73FF,	CF,	AMD Radeon RX 6600 LE
+73FF,	DF,	AMD Radeon RX 6750 GRE 10GB
+7408,	00,	AMD Instinct MI250X
+740C,	01,	AMD Instinct MI250X / MI250
+740F,	02,	AMD Instinct MI210
+7421,	00,	AMD Radeon Pro W6500M
+7422,	00,	AMD Radeon Pro W6400
+7423,	00,	AMD Radeon Pro W6300M
+7423,	01,	AMD Radeon Pro W6300
+7424,	00,	AMD Radeon RX 6300
+743F,	C1,	AMD Radeon RX 6500 XT
+743F,	C3,	AMD Radeon RX 6500
+743F,	C3,	AMD Radeon RX 6500M
+743F,	C7,	AMD Radeon RX 6400
+743F,	C8,	AMD Radeon RX 6500M
+743F,	CC,	AMD Radeon 6550S
+743F,	CE,	AMD Radeon RX 6450M
+743F,	CF,	AMD Radeon RX 6300M
+743F,	D3,	AMD Radeon RX 6550M
+743F,	D7,	AMD Radeon RX 6400
+7448,	00,	AMD Radeon Pro W7900
+7449,	00,	AMD Radeon Pro W7800 48GB
+744A,	00,	AMD Radeon Pro W7900 Dual Slot
+744B,	00,	AMD Radeon Pro W7900D
+744C,	C8,	AMD Radeon RX 7900 XTX
+744C,	CC,	AMD Radeon RX 7900 XT
+744C,	CE,	AMD Radeon RX 7900 GRE
+744C,	CF,	AMD Radeon RX 7900M
+745E,	CC,	AMD Radeon Pro W7800
+7460,	00,	AMD Radeon Pro V710
+7461,	00,	AMD Radeon Pro V710 MxGPU
+7470,	00,	AMD Radeon Pro W7700
+747E,	C8,	AMD Radeon RX 7800 XT
+747E,	D8,	AMD Radeon RX 7800M
+747E,	DB,	AMD Radeon RX 7700
+747E,	FF,	AMD Radeon RX 7700 XT
+7480,	00,	AMD Radeon Pro W7600
+7480,	C0,	AMD Radeon RX 7600 XT
+7480,	C1,	AMD Radeon RX 7700S
+7480,	C2,	AMD Radeon RX 7650 GRE
+7480,	C3,	AMD Radeon RX 7600S
+7480,	C7,	AMD Radeon RX 7600M XT
+7480,	CF,	AMD Radeon RX 7600
+7481,   C7,     AMD Steam Machine
+7483,	CF,	AMD Radeon RX 7600M
+7489,	00,	AMD Radeon Pro W7500
+7499,	00,	AMD Radeon Pro W7400
+7499,	C0,	AMD Radeon RX 7400
+7499,	C1,	AMD Radeon RX 7300
+74A0,	00,	AMD Instinct MI300A
+74A1,	00,	AMD Instinct MI300X
+74A2,	00,	AMD Instinct MI308X
+74A5,	00,	AMD Instinct MI325X
+74A8,	00,	AMD Instinct MI308X HF
+74A9,	00,	AMD Instinct MI300X HF
+74B5,	00,	AMD Instinct MI300X VF
+74B6,	00,	AMD Instinct MI308X
+74BD,	00,	AMD Instinct MI300X HF
+7550,	C0,	AMD Radeon RX 9070 XT
+7550,	C2,	AMD Radeon RX 9070 GRE
+7550,	C3,	AMD Radeon RX 9070
+7551,	C0,	AMD Radeon AI PRO R9700
+7590,	C0,	AMD Radeon RX 9060 XT
+7590,	C7,	AMD Radeon RX 9060
+75A0,	C0,	AMD Instinct MI350X
+75A3,	C0,	AMD Instinct MI355X
+75B0,	C0,	AMD Instinct MI350X VF
+75B3,	C0,	AMD Instinct MI355X VF
+9830,	00,	AMD Radeon HD 8400 / R3 Series
+9831,	00,	AMD Radeon HD 8400E
+9832,	00,	AMD Radeon HD 8330
+9833,	00,	AMD Radeon HD 8330E
+9834,	00,	AMD Radeon HD 8210
+9835,	00,	AMD Radeon HD 8210E
+9836,	00,	AMD Radeon HD 8200 / R3 Series
+9837,	00,	AMD Radeon HD 8280E
+9838,	00,	AMD Radeon HD 8200 / R3 series
+9839,	00,	AMD Radeon HD 8180
+983D,	00,	AMD Radeon HD 8250
+9850,	00,	AMD Radeon R3 Graphics
+9850,	03,	AMD Radeon R3 Graphics
+9850,	40,	AMD Radeon R2 Graphics
+9850,	45,	AMD Radeon R3 Graphics
+9851,	00,	AMD Radeon R4 Graphics
+9851,	01,	AMD Radeon R5E Graphics
+9851,	05,	AMD Radeon R5 Graphics
+9851,	06,	AMD Radeon R5E Graphics
+9851,	40,	AMD Radeon R4 Graphics
+9851,	45,	AMD Radeon R5 Graphics
+9852,	00,	AMD Radeon R2 Graphics
+9852,	40,	AMD Radeon E1 Graphics
+9853,	00,	AMD Radeon R2 Graphics
+9853,	01,	AMD Radeon R4E Graphics
+9853,	03,	AMD Radeon R2 Graphics
+9853,	05,	AMD Radeon R1E Graphics
+9853,	06,	AMD Radeon R1E Graphics
+9853,	07,	AMD Radeon R1E Graphics
+9853,	08,	AMD Radeon R1E Graphics
+9853,	40,	AMD Radeon R2 Graphics
+9854,	00,	AMD Radeon R3 Graphics
+9854,	01,	AMD Radeon R3E Graphics
+9854,	02,	AMD Radeon R3 Graphics
+9854,	05,	AMD Radeon R2 Graphics
+9854,	06,	AMD Radeon R4 Graphics
+9854,	07,	AMD Radeon R3 Graphics
+9855,	02,	AMD Radeon R6 Graphics
+9855,	05,	AMD Radeon R4 Graphics
+9856,	00,	AMD Radeon R2 Graphics
+9856,	01,	AMD Radeon R2E Graphics
+9856,	02,	AMD Radeon R2 Graphics
+9856,	05,	AMD Radeon R1E Graphics
+9856,	06,	AMD Radeon R2 Graphics
+9856,	07,	AMD Radeon R1E Graphics
+9856,	08,	AMD Radeon R1E Graphics
+9856,	13,	AMD Radeon R1E Graphics
+9874,	81,	AMD Radeon R6 Graphics
+9874,	84,	AMD Radeon R7 Graphics
+9874,	85,	AMD Radeon R6 Graphics
+9874,	87,	AMD Radeon R5 Graphics
+9874,	88,	AMD Radeon R7E Graphics
+9874,	89,	AMD Radeon R6E Graphics
+9874,	C4,	AMD Radeon R7 Graphics
+9874,	C5,	AMD Radeon R6 Graphics
+9874,	C6,	AMD Radeon R6 Graphics
+9874,	C7,	AMD Radeon R5 Graphics
+9874,	C8,	AMD Radeon R7 Graphics
+9874,	C9,	AMD Radeon R7 Graphics
+9874,	CA,	AMD Radeon R5 Graphics
+9874,	CB,	AMD Radeon R5 Graphics
+9874,	CC,	AMD Radeon R7 Graphics
+9874,	CD,	AMD Radeon R7 Graphics
+9874,	CE,	AMD Radeon R5 Graphics
+9874,	E1,	AMD Radeon R7 Graphics
+9874,	E2,	AMD Radeon R7 Graphics
+9874,	E3,	AMD Radeon R7 Graphics
+9874,	E4,	AMD Radeon R7 Graphics
+9874,	E5,	AMD Radeon R5 Graphics
+9874,	E6,	AMD Radeon R5 Graphics
+98E4,	80,	AMD Radeon R5E Graphics
+98E4,	81,	AMD Radeon R4E Graphics
+98E4,	83,	AMD Radeon R2E Graphics
+98E4,	84,	AMD Radeon R2E Graphics
+98E4,	86,	AMD Radeon R1E Graphics
+98E4,	C0,	AMD Radeon R4 Graphics
+98E4,	C1,	AMD Radeon R5 Graphics
+98E4,	C2,	AMD Radeon R4 Graphics
+98E4,	C4,	AMD Radeon R5 Graphics
+98E4,	C6,	AMD Radeon R5 Graphics
+98E4,	C8,	AMD Radeon R4 Graphics
+98E4,	C9,	AMD Radeon R4 Graphics
+98E4,	CA,	AMD Radeon R5 Graphics
+98E4,	D0,	AMD Radeon R2 Graphics
+98E4,	D1,	AMD Radeon R2 Graphics
+98E4,	D2,	AMD Radeon R2 Graphics
+98E4,	D4,	AMD Radeon R2 Graphics
+98E4,	D9,	AMD Radeon R5 Graphics
+98E4,	DA,	AMD Radeon R5 Graphics
+98E4,	DB,	AMD Radeon R3 Graphics
+98E4,	E1,	AMD Radeon R3 Graphics
+98E4,	E2,	AMD Radeon R3 Graphics
+98E4,	E9,	AMD Radeon R4 Graphics
+98E4,	EA,	AMD Radeon R4 Graphics
+98E4,	EB,	AMD Radeon R3 Graphics
+98E4,	EB,	AMD Radeon R4 Graphics

agent/test-data/nvtop.json

@@ -0,0 +1,34 @@
+[
+  {
+   "device_name": "NVIDIA GeForce RTX 3050 Ti Laptop GPU",
+   "gpu_clock": "1485MHz",
+   "mem_clock": "6001MHz",
+   "temp": "48C",
+   "fan_speed": null,
+   "power_draw": "13W",
+   "gpu_util": "5%",
+   "encode": "0%",
+   "decode": "0%",
+   "mem_util": "8%",
+   "mem_total": "4294967296",
+   "mem_used": "349372416",
+   "mem_free": "3945594880",
+   "processes" : []
+  },
+  {
+   "device_name": "AMD Radeon 680M",
+   "gpu_clock": "2200MHz",
+   "mem_clock": "2400MHz",
+   "temp": "48C",
+   "fan_speed": "CPU Fan",
+   "power_draw": "9W",
+   "gpu_util": "12%",
+   "encode": null,
+   "decode": "0%",
+   "mem_util": "7%",
+   "mem_total": "16929173504",
+   "mem_used": "1213784064",
+   "mem_free": "15715389440",
+   "processes" : []
+  }
+]

internal/alerts/alerts_smart.go

@@ -2,18 +2,18 @@ package alerts
 
 import (
 	"fmt"
+	"strings"
 
 	"github.com/pocketbase/pocketbase/core"
 )
 
-// handleSmartDeviceAlert sends alerts when a SMART device state changes from PASSED to FAILED.
+// handleSmartDeviceAlert sends alerts when a SMART device state worsens into WARNING/FAILED.
 // This is automatic and does not require user opt-in.
 func (am *AlertManager) handleSmartDeviceAlert(e *core.RecordEvent) error {
 	oldState := e.Record.Original().GetString("state")
 	newState := e.Record.GetString("state")
 
-	// Only alert when transitioning from PASSED to FAILED
-	if oldState != "PASSED" || newState != "FAILED" {
+	if !shouldSendSmartDeviceAlert(oldState, newState) {
 		return e.Next()
 	}
 
@@ -32,14 +32,15 @@ func (am *AlertManager) handleSmartDeviceAlert(e *core.RecordEvent) error {
 	systemName := systemRecord.GetString("name")
 	deviceName := e.Record.GetString("name")
 	model := e.Record.GetString("model")
+	statusLabel := smartStateLabel(newState)
 
 	// Build alert message
-	title := fmt.Sprintf("SMART failure on %s: %s \U0001F534", systemName, deviceName)
+	title := fmt.Sprintf("SMART %s on %s: %s %s", statusLabel, systemName, deviceName, smartStateEmoji(newState))
 	var message string
 	if model != "" {
-		message = fmt.Sprintf("Disk %s (%s) SMART status changed to FAILED", deviceName, model)
+		message = fmt.Sprintf("Disk %s (%s) SMART status changed to %s", deviceName, model, newState)
 	} else {
-		message = fmt.Sprintf("Disk %s SMART status changed to FAILED", deviceName)
+		message = fmt.Sprintf("Disk %s SMART status changed to %s", deviceName, newState)
 	}
 
 	// Get users associated with the system
@@ -65,3 +66,42 @@ func (am *AlertManager) handleSmartDeviceAlert(e *core.RecordEvent) error {
 	return e.Next()
 }
 
+func shouldSendSmartDeviceAlert(oldState, newState string) bool {
+	oldSeverity := smartStateSeverity(oldState)
+	newSeverity := smartStateSeverity(newState)
+
+	// Ignore unknown states and recoveries; only alert on worsening transitions
+	// from known-good/degraded states into WARNING/FAILED.
+	return oldSeverity >= 1 && newSeverity > oldSeverity
+}
+
+func smartStateSeverity(state string) int {
+	switch state {
+	case "PASSED":
+		return 1
+	case "WARNING":
+		return 2
+	case "FAILED":
+		return 3
+	default:
+		return 0
+	}
+}
+
+func smartStateEmoji(state string) string {
+	switch state {
+	case "WARNING":
+		return "\U0001F7E0"
+	default:
+		return "\U0001F534"
+	}
+}
+
+func smartStateLabel(state string) string {
+	switch state {
+	case "FAILED":
+		return "failure"
+	default:
+		return strings.ToLower(state)
+	}
+}

internal/alerts/alerts_smart_test.go

@@ -58,6 +58,74 @@ func TestSmartDeviceAlert(t *testing.T) {
 	assert.Contains(t, lastMessage.Text, "FAILED")
 }
 
+func TestSmartDeviceAlertPassedToWarning(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	smartDevice, err := beszelTests.CreateRecord(hub, "smart_devices", map[string]any{
+		"system": system.Id,
+		"name":   "/dev/mmcblk0",
+		"model":  "eMMC",
+		"state":  "PASSED",
+	})
+	assert.NoError(t, err)
+
+	smartDevice, err = hub.FindRecordById("smart_devices", smartDevice.Id)
+	assert.NoError(t, err)
+
+	smartDevice.Set("state", "WARNING")
+	err = hub.Save(smartDevice)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	assert.EqualValues(t, 1, hub.TestMailer.TotalSend(), "should have 1 email sent after state changed to WARNING")
+	lastMessage := hub.TestMailer.LastMessage()
+	assert.Contains(t, lastMessage.Subject, "SMART warning on test-system")
+	assert.Contains(t, lastMessage.Text, "WARNING")
+}
+
+func TestSmartDeviceAlertWarningToFailed(t *testing.T) {
+	hub, user := beszelTests.GetHubWithUser(t)
+	defer hub.Cleanup()
+
+	system, err := beszelTests.CreateRecord(hub, "systems", map[string]any{
+		"name":  "test-system",
+		"users": []string{user.Id},
+		"host":  "127.0.0.1",
+	})
+	assert.NoError(t, err)
+
+	smartDevice, err := beszelTests.CreateRecord(hub, "smart_devices", map[string]any{
+		"system": system.Id,
+		"name":   "/dev/mmcblk0",
+		"model":  "eMMC",
+		"state":  "WARNING",
+	})
+	assert.NoError(t, err)
+
+	smartDevice, err = hub.FindRecordById("smart_devices", smartDevice.Id)
+	assert.NoError(t, err)
+
+	smartDevice.Set("state", "FAILED")
+	err = hub.Save(smartDevice)
+	assert.NoError(t, err)
+
+	time.Sleep(50 * time.Millisecond)
+
+	assert.EqualValues(t, 1, hub.TestMailer.TotalSend(), "should have 1 email sent after state changed from WARNING to FAILED")
+	lastMessage := hub.TestMailer.LastMessage()
+	assert.Contains(t, lastMessage.Subject, "SMART failure on test-system")
+	assert.Contains(t, lastMessage.Text, "FAILED")
+}
+
 func TestSmartDeviceAlertNoAlertOnNonPassedToFailed(t *testing.T) {
 	hub, user := beszelTests.GetHubWithUser(t)
 	defer hub.Cleanup()
@@ -83,7 +151,8 @@ func TestSmartDeviceAlertNoAlertOnNonPassedToFailed(t *testing.T) {
 	smartDevice, err = hub.FindRecordById("smart_devices", smartDevice.Id)
 	assert.NoError(t, err)
 
-	// Update the state from UNKNOWN to FAILED - should NOT trigger alert
+	// Update the state from UNKNOWN to FAILED - should NOT trigger alert.
+	// We only alert from known healthy/degraded states.
 	smartDevice.Set("state", "FAILED")
 	err = hub.Save(smartDevice)
 	assert.NoError(t, err)

internal/dockerfile_agent

@@ -23,6 +23,9 @@ COPY --from=builder /agent /agent
 # this is so we don't need to create the /tmp directory in the scratch container
 COPY --from=builder /tmp /tmp
 
+# AMD GPU name lookup (used by agent on Linux when /usr/share/libdrm/amdgpu.ids is read)
+COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
+
 # Ensure data persistence across container recreations
 VOLUME ["/var/lib/beszel-agent"]
 

internal/dockerfile_agent_alpine

@@ -20,6 +20,9 @@ RUN rm -rf /tmp/*
 FROM alpine:3.23
 COPY --from=builder /agent /agent
 
+# AMD GPU name lookup (used by agent on Linux when /usr/share/libdrm/amdgpu.ids is read)
+COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
+
 RUN apk add --no-cache smartmontools
 
 # Ensure data persistence across container recreations

internal/dockerfile_agent_nvidia

@@ -37,6 +37,9 @@ RUN apt-get update && apt-get install -y \
 FROM nvidia/cuda:12.2.2-base-ubuntu22.04
 COPY --from=builder /agent /agent
 
+# AMD GPU name lookup (used by agent on hybrid laptops when /usr/share/libdrm/amdgpu.ids is read)
+COPY --from=builder /app/agent/test-data/amdgpu.ids /usr/share/libdrm/amdgpu.ids
+
 # Copy smartmontools binaries and config files
 COPY --from=smartmontools-builder /usr/sbin/smartctl /usr/sbin/smartctl
 

internal/hub/heartbeat/heartbeat.go

@@ -0,0 +1,303 @@
+// Package heartbeat sends periodic outbound pings to an external monitoring
+// endpoint (e.g. BetterStack, Uptime Kuma, Healthchecks.io) so operators can
+// monitor Beszel without exposing it to the internet.
+package heartbeat
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/henrygd/beszel"
+	"github.com/pocketbase/pocketbase/core"
+)
+
+// Default values for heartbeat configuration.
+const (
+	defaultInterval = 60 // seconds
+	httpTimeout     = 10 * time.Second
+)
+
+// Payload is the JSON body sent with each heartbeat request.
+type Payload struct {
+	// Status is "ok" when all non-paused systems are up, "warn" when alerts
+	// are triggered but no systems are down, and "error" when any system is down.
+	Status    string         `json:"status"`
+	Timestamp string         `json:"timestamp"`
+	Msg       string         `json:"msg"`
+	Systems   SystemsSummary `json:"systems"`
+	Down      []SystemInfo   `json:"down_systems,omitempty"`
+	Alerts    []AlertInfo    `json:"triggered_alerts,omitempty"`
+	Version   string         `json:"beszel_version"`
+}
+
+// SystemsSummary contains counts of systems by status.
+type SystemsSummary struct {
+	Total   int `json:"total"`
+	Up      int `json:"up"`
+	Down    int `json:"down"`
+	Paused  int `json:"paused"`
+	Pending int `json:"pending"`
+}
+
+// SystemInfo identifies a system that is currently down.
+type SystemInfo struct {
+	ID   string `json:"id" db:"id"`
+	Name string `json:"name" db:"name"`
+	Host string `json:"host" db:"host"`
+}
+
+// AlertInfo describes a currently triggered alert.
+type AlertInfo struct {
+	SystemID   string  `json:"system_id"`
+	SystemName string  `json:"system_name"`
+	AlertName  string  `json:"alert_name"`
+	Threshold  float64 `json:"threshold"`
+}
+
+// Config holds heartbeat settings read from environment variables.
+type Config struct {
+	URL      string // endpoint to ping
+	Interval int    // seconds between pings
+	Method   string // HTTP method (GET or POST, default POST)
+}
+
+// Heartbeat manages the periodic outbound health check.
+type Heartbeat struct {
+	app    core.App
+	config Config
+	client *http.Client
+}
+
+// New creates a Heartbeat if configuration is present.
+// Returns nil if HEARTBEAT_URL is not set (feature disabled).
+func New(app core.App, getEnv func(string) (string, bool)) *Heartbeat {
+	url, _ := getEnv("HEARTBEAT_URL")
+	url = strings.TrimSpace(url)
+	if app == nil || url == "" {
+		return nil
+	}
+
+	interval := defaultInterval
+	if v, ok := getEnv("HEARTBEAT_INTERVAL"); ok {
+		if parsed, err := strconv.Atoi(v); err == nil && parsed > 0 {
+			interval = parsed
+		}
+	}
+
+	method := http.MethodPost
+	if v, ok := getEnv("HEARTBEAT_METHOD"); ok {
+		v = strings.ToUpper(strings.TrimSpace(v))
+		if v == http.MethodGet || v == http.MethodHead {
+			method = v
+		}
+	}
+
+	return &Heartbeat{
+		app: app,
+		config: Config{
+			URL:      url,
+			Interval: interval,
+			Method:   method,
+		},
+		client: &http.Client{Timeout: httpTimeout},
+	}
+}
+
+// Start begins the heartbeat loop. It blocks and should be called in a goroutine.
+// The loop runs until the provided stop channel is closed.
+func (hb *Heartbeat) Start(stop <-chan struct{}) {
+	sanitizedURL := sanitizeHeartbeatURL(hb.config.URL)
+	hb.app.Logger().Info("Heartbeat enabled",
+		"url", sanitizedURL,
+		"interval", fmt.Sprintf("%ds", hb.config.Interval),
+		"method", hb.config.Method,
+	)
+
+	// Send an initial heartbeat immediately on startup.
+	hb.send()
+
+	ticker := time.NewTicker(time.Duration(hb.config.Interval) * time.Second)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-stop:
+			return
+		case <-ticker.C:
+			hb.send()
+		}
+	}
+}
+
+// Send performs a single heartbeat ping. Exposed for the test-heartbeat API endpoint.
+func (hb *Heartbeat) Send() error {
+	return hb.send()
+}
+
+// GetConfig returns the current heartbeat configuration.
+func (hb *Heartbeat) GetConfig() Config {
+	return hb.config
+}
+
+func (hb *Heartbeat) send() error {
+	var req *http.Request
+	var err error
+	method := normalizeMethod(hb.config.Method)
+
+	if method == http.MethodGet || method == http.MethodHead {
+		req, err = http.NewRequest(method, hb.config.URL, nil)
+	} else {
+		payload, payloadErr := hb.buildPayload()
+		if payloadErr != nil {
+			hb.app.Logger().Error("Heartbeat: failed to build payload", "err", payloadErr)
+			return payloadErr
+		}
+
+		body, jsonErr := json.Marshal(payload)
+		if jsonErr != nil {
+			hb.app.Logger().Error("Heartbeat: failed to marshal payload", "err", jsonErr)
+			return jsonErr
+		}
+		req, err = http.NewRequest(http.MethodPost, hb.config.URL, bytes.NewReader(body))
+		if err == nil {
+			req.Header.Set("Content-Type", "application/json")
+		}
+	}
+
+	if err != nil {
+		hb.app.Logger().Error("Heartbeat: failed to create request", "err", err)
+		return err
+	}
+
+	req.Header.Set("User-Agent", "Beszel-Heartbeat")
+
+	resp, err := hb.client.Do(req)
+	if err != nil {
+		hb.app.Logger().Error("Heartbeat: request failed", "url", sanitizeHeartbeatURL(hb.config.URL), "err", err)
+		return err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode >= 400 {
+		hb.app.Logger().Warn("Heartbeat: non-success response",
+			"url", sanitizeHeartbeatURL(hb.config.URL),
+			"status", resp.StatusCode,
+		)
+		return fmt.Errorf("heartbeat endpoint returned status %d", resp.StatusCode)
+	}
+
+	return nil
+}
+
+func (hb *Heartbeat) buildPayload() (*Payload, error) {
+	db := hb.app.DB()
+
+	// Count systems by status.
+	var systemCounts []struct {
+		Status string `db:"status"`
+		Count  int    `db:"cnt"`
+	}
+	err := db.NewQuery("SELECT status, COUNT(*) as cnt FROM systems GROUP BY status").All(&systemCounts)
+	if err != nil {
+		return nil, fmt.Errorf("query system counts: %w", err)
+	}
+
+	summary := SystemsSummary{}
+	for _, sc := range systemCounts {
+		switch sc.Status {
+		case "up":
+			summary.Up = sc.Count
+		case "down":
+			summary.Down = sc.Count
+		case "paused":
+			summary.Paused = sc.Count
+		case "pending":
+			summary.Pending = sc.Count
+		}
+		summary.Total += sc.Count
+	}
+
+	// Get names of down systems.
+	var downSystems []SystemInfo
+	if summary.Down > 0 {
+		err = db.NewQuery("SELECT id, name, host FROM systems WHERE status = 'down'").All(&downSystems)
+		if err != nil {
+			return nil, fmt.Errorf("query down systems: %w", err)
+		}
+	}
+
+	// Get triggered alerts with system names.
+	var triggeredAlerts []struct {
+		SystemID   string  `db:"system"`
+		SystemName string  `db:"system_name"`
+		AlertName  string  `db:"name"`
+		Value      float64 `db:"value"`
+	}
+	err = db.NewQuery(`
+		SELECT a.system, s.name as system_name, a.name, a.value
+		FROM alerts a
+		JOIN systems s ON a.system = s.id
+		WHERE a.triggered = true
+	`).All(&triggeredAlerts)
+	if err != nil {
+		// Non-fatal: alerts info is supplementary.
+		triggeredAlerts = nil
+	}
+
+	alerts := make([]AlertInfo, 0, len(triggeredAlerts))
+	for _, ta := range triggeredAlerts {
+		alerts = append(alerts, AlertInfo{
+			SystemID:   ta.SystemID,
+			SystemName: ta.SystemName,
+			AlertName:  ta.AlertName,
+			Threshold:  ta.Value,
+		})
+	}
+
+	// Determine overall status.
+	status := "ok"
+	msg := "All systems operational"
+	if summary.Down > 0 {
+		status = "error"
+		names := make([]string, len(downSystems))
+		for i, ds := range downSystems {
+			names[i] = ds.Name
+		}
+		msg = fmt.Sprintf("%d system(s) down: %s", summary.Down, strings.Join(names, ", "))
+	} else if len(alerts) > 0 {
+		status = "warn"
+		msg = fmt.Sprintf("%d alert(s) triggered", len(alerts))
+	}
+
+	return &Payload{
+		Status:    status,
+		Timestamp: time.Now().UTC().Format(time.RFC3339),
+		Msg:       msg,
+		Systems:   summary,
+		Down:      downSystems,
+		Alerts:    alerts,
+		Version:   beszel.Version,
+	}, nil
+}
+
+func normalizeMethod(method string) string {
+	upper := strings.ToUpper(strings.TrimSpace(method))
+	if upper == http.MethodGet || upper == http.MethodHead || upper == http.MethodPost {
+		return upper
+	}
+	return http.MethodPost
+}
+
+func sanitizeHeartbeatURL(rawURL string) string {
+	parsed, err := url.Parse(strings.TrimSpace(rawURL))
+	if err != nil || parsed.Scheme == "" || parsed.Host == "" {
+		return "<invalid-url>"
+	}
+	return parsed.Scheme + "://" + parsed.Host
+}

internal/hub/heartbeat/heartbeat_test.go

@@ -0,0 +1,258 @@
+//go:build testing
+// +build testing
+
+package heartbeat_test
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/henrygd/beszel/internal/hub/heartbeat"
+	beszeltests "github.com/henrygd/beszel/internal/tests"
+	"github.com/pocketbase/pocketbase/core"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestNew(t *testing.T) {
+	t.Run("returns nil when app is missing", func(t *testing.T) {
+		hb := heartbeat.New(nil, envGetter(map[string]string{
+			"HEARTBEAT_URL": "https://heartbeat.example.com/ping",
+		}))
+		assert.Nil(t, hb)
+	})
+
+	t.Run("returns nil when URL is missing", func(t *testing.T) {
+		app := newTestHub(t)
+		hb := heartbeat.New(app.App, func(string) (string, bool) {
+			return "", false
+		})
+		assert.Nil(t, hb)
+	})
+
+	t.Run("parses and normalizes config values", func(t *testing.T) {
+		app := newTestHub(t)
+		env := map[string]string{
+			"HEARTBEAT_URL":      "  https://heartbeat.example.com/ping  ",
+			"HEARTBEAT_INTERVAL": "90",
+			"HEARTBEAT_METHOD":   "head",
+		}
+		getEnv := func(key string) (string, bool) {
+			v, ok := env[key]
+			return v, ok
+		}
+
+		hb := heartbeat.New(app.App, getEnv)
+		require.NotNil(t, hb)
+		cfg := hb.GetConfig()
+		assert.Equal(t, "https://heartbeat.example.com/ping", cfg.URL)
+		assert.Equal(t, 90, cfg.Interval)
+		assert.Equal(t, http.MethodHead, cfg.Method)
+	})
+}
+
+func TestSendGETDoesNotRequireAppOrDB(t *testing.T) {
+	app := newTestHub(t)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		assert.Equal(t, http.MethodGet, r.Method)
+		assert.Equal(t, "Beszel-Heartbeat", r.Header.Get("User-Agent"))
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer server.Close()
+
+	hb := heartbeat.New(app.App, envGetter(map[string]string{
+		"HEARTBEAT_URL":    server.URL,
+		"HEARTBEAT_METHOD": "GET",
+	}))
+	require.NotNil(t, hb)
+
+	require.NoError(t, hb.Send())
+}
+
+func TestSendReturnsErrorOnHTTPFailureStatus(t *testing.T) {
+	app := newTestHub(t)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusInternalServerError)
+	}))
+	defer server.Close()
+
+	hb := heartbeat.New(app.App, envGetter(map[string]string{
+		"HEARTBEAT_URL":    server.URL,
+		"HEARTBEAT_METHOD": "GET",
+	}))
+	require.NotNil(t, hb)
+
+	err := hb.Send()
+	require.Error(t, err)
+	assert.ErrorContains(t, err, "heartbeat endpoint returned status 500")
+}
+
+func TestSendPOSTBuildsExpectedStatuses(t *testing.T) {
+	tests := []struct {
+		name           string
+		setup          func(t *testing.T, app *beszeltests.TestHub, user *core.Record)
+		expectStatus   string
+		expectMsgPart  string
+		expectDown     int
+		expectAlerts   int
+		expectTotal    int
+		expectUp       int
+		expectPaused   int
+		expectPending  int
+		expectDownSumm int
+	}{
+		{
+			name: "error when at least one system is down",
+			setup: func(t *testing.T, app *beszeltests.TestHub, user *core.Record) {
+				downSystem := createTestSystem(t, app, user.Id, "db-1", "10.0.0.1", "down")
+				_ = createTestSystem(t, app, user.Id, "web-1", "10.0.0.2", "up")
+				createTriggeredAlert(t, app, user.Id, downSystem.Id, "CPU", 95)
+			},
+			expectStatus:   "error",
+			expectMsgPart:  "1 system(s) down",
+			expectDown:     1,
+			expectAlerts:   1,
+			expectTotal:    2,
+			expectUp:       1,
+			expectDownSumm: 1,
+		},
+		{
+			name: "warn when only alerts are triggered",
+			setup: func(t *testing.T, app *beszeltests.TestHub, user *core.Record) {
+				system := createTestSystem(t, app, user.Id, "api-1", "10.1.0.1", "up")
+				createTriggeredAlert(t, app, user.Id, system.Id, "CPU", 90)
+			},
+			expectStatus:   "warn",
+			expectMsgPart:  "1 alert(s) triggered",
+			expectDown:     0,
+			expectAlerts:   1,
+			expectTotal:    1,
+			expectUp:       1,
+			expectDownSumm: 0,
+		},
+		{
+			name: "ok when no down systems and no alerts",
+			setup: func(t *testing.T, app *beszeltests.TestHub, user *core.Record) {
+				_ = createTestSystem(t, app, user.Id, "node-1", "10.2.0.1", "up")
+				_ = createTestSystem(t, app, user.Id, "node-2", "10.2.0.2", "paused")
+				_ = createTestSystem(t, app, user.Id, "node-3", "10.2.0.3", "pending")
+			},
+			expectStatus:   "ok",
+			expectMsgPart:  "All systems operational",
+			expectDown:     0,
+			expectAlerts:   0,
+			expectTotal:    3,
+			expectUp:       1,
+			expectPaused:   1,
+			expectPending:  1,
+			expectDownSumm: 0,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			app := newTestHub(t)
+			user := createTestUser(t, app)
+			tt.setup(t, app, user)
+
+			type requestCapture struct {
+				method      string
+				userAgent   string
+				contentType string
+				payload     heartbeat.Payload
+			}
+
+			captured := make(chan requestCapture, 1)
+			server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				defer r.Body.Close()
+				body, err := io.ReadAll(r.Body)
+				require.NoError(t, err)
+
+				var payload heartbeat.Payload
+				require.NoError(t, json.Unmarshal(body, &payload))
+				captured <- requestCapture{
+					method:      r.Method,
+					userAgent:   r.Header.Get("User-Agent"),
+					contentType: r.Header.Get("Content-Type"),
+					payload:     payload,
+				}
+				w.WriteHeader(http.StatusNoContent)
+			}))
+			defer server.Close()
+
+			hb := heartbeat.New(app.App, envGetter(map[string]string{
+				"HEARTBEAT_URL":    server.URL,
+				"HEARTBEAT_METHOD": "POST",
+			}))
+			require.NotNil(t, hb)
+			require.NoError(t, hb.Send())
+
+			req := <-captured
+			assert.Equal(t, http.MethodPost, req.method)
+			assert.Equal(t, "Beszel-Heartbeat", req.userAgent)
+			assert.Equal(t, "application/json", req.contentType)
+
+			assert.Equal(t, tt.expectStatus, req.payload.Status)
+			assert.Contains(t, req.payload.Msg, tt.expectMsgPart)
+			assert.Equal(t, tt.expectDown, len(req.payload.Down))
+			assert.Equal(t, tt.expectAlerts, len(req.payload.Alerts))
+			assert.Equal(t, tt.expectTotal, req.payload.Systems.Total)
+			assert.Equal(t, tt.expectUp, req.payload.Systems.Up)
+			assert.Equal(t, tt.expectDownSumm, req.payload.Systems.Down)
+			assert.Equal(t, tt.expectPaused, req.payload.Systems.Paused)
+			assert.Equal(t, tt.expectPending, req.payload.Systems.Pending)
+		})
+	}
+}
+
+func newTestHub(t *testing.T) *beszeltests.TestHub {
+	t.Helper()
+	app, err := beszeltests.NewTestHub(t.TempDir())
+	require.NoError(t, err)
+	t.Cleanup(app.Cleanup)
+	return app
+}
+
+func createTestUser(t *testing.T, app *beszeltests.TestHub) *core.Record {
+	t.Helper()
+	user, err := beszeltests.CreateUser(app.App, "admin@example.com", "password123")
+	require.NoError(t, err)
+	return user
+}
+
+func createTestSystem(t *testing.T, app *beszeltests.TestHub, userID, name, host, status string) *core.Record {
+	t.Helper()
+	system, err := beszeltests.CreateRecord(app.App, "systems", map[string]any{
+		"name":   name,
+		"host":   host,
+		"port":   "45876",
+		"users":  []string{userID},
+		"status": status,
+	})
+	require.NoError(t, err)
+	return system
+}
+
+func createTriggeredAlert(t *testing.T, app *beszeltests.TestHub, userID, systemID, name string, threshold float64) *core.Record {
+	t.Helper()
+	alert, err := beszeltests.CreateRecord(app.App, "alerts", map[string]any{
+		"name":      name,
+		"system":    systemID,
+		"user":      userID,
+		"value":     threshold,
+		"min":       0,
+		"triggered": true,
+	})
+	require.NoError(t, err)
+	return alert
+}
+
+func envGetter(values map[string]string) func(string) (string, bool) {
+	return func(key string) (string, bool) {
+		v, ok := values[key]
+		return v, ok
+	}
+}

internal/hub/hub.go

@@ -15,6 +15,7 @@ import (
 	"github.com/henrygd/beszel"
 	"github.com/henrygd/beszel/internal/alerts"
 	"github.com/henrygd/beszel/internal/hub/config"
+	"github.com/henrygd/beszel/internal/hub/heartbeat"
 	"github.com/henrygd/beszel/internal/hub/systems"
 	"github.com/henrygd/beszel/internal/records"
 	"github.com/henrygd/beszel/internal/users"
@@ -33,6 +34,8 @@ type Hub struct {
 	um     *users.UserManager
 	rm     *records.RecordManager
 	sm     *systems.SystemManager
+	hb     *heartbeat.Heartbeat
+	hbStop chan struct{}
 	pubKey string
 	signer ssh.Signer
 	appURL string
@@ -48,6 +51,10 @@ func NewHub(app core.App) *Hub {
 	hub.rm = records.NewRecordManager(hub)
 	hub.sm = systems.NewSystemManager(hub)
 	hub.appURL, _ = GetEnv("APP_URL")
+	hub.hb = heartbeat.New(app, GetEnv)
+	if hub.hb != nil {
+		hub.hbStop = make(chan struct{})
+	}
 	return hub
 }
 
@@ -88,6 +95,10 @@ func (h *Hub) StartHub() error {
 		if err := h.sm.Initialize(); err != nil {
 			return err
 		}
+		// start heartbeat if configured
+		if h.hb != nil {
+			go h.hb.Start(h.hbStop)
+		}
 		return e.Next()
 	})
 
@@ -287,6 +298,9 @@ func (h *Hub) registerApiRoutes(se *core.ServeEvent) error {
 	})
 	// send test notification
 	apiAuth.POST("/test-notification", h.SendTestNotification)
+	// heartbeat status and test
+	apiAuth.GET("/heartbeat-status", h.getHeartbeatStatus)
+	apiAuth.POST("/test-heartbeat", h.testHeartbeat)
 	// get config.yml content
 	apiAuth.GET("/config-yaml", config.GetYamlConfig)
 	// handle agent websocket connection
@@ -403,6 +417,42 @@ func (h *Hub) getUniversalToken(e *core.RequestEvent) error {
 	return e.JSON(http.StatusOK, response)
 }
 
+// getHeartbeatStatus returns current heartbeat configuration and whether it's enabled
+func (h *Hub) getHeartbeatStatus(e *core.RequestEvent) error {
+	if e.Auth.GetString("role") != "admin" {
+		return e.ForbiddenError("Requires admin role", nil)
+	}
+	if h.hb == nil {
+		return e.JSON(http.StatusOK, map[string]any{
+			"enabled": false,
+			"msg":     "Set HEARTBEAT_URL to enable outbound heartbeat monitoring",
+		})
+	}
+	cfg := h.hb.GetConfig()
+	return e.JSON(http.StatusOK, map[string]any{
+		"enabled":  true,
+		"url":      cfg.URL,
+		"interval": cfg.Interval,
+		"method":   cfg.Method,
+	})
+}
+
+// testHeartbeat triggers a single heartbeat ping and returns the result
+func (h *Hub) testHeartbeat(e *core.RequestEvent) error {
+	if e.Auth.GetString("role") != "admin" {
+		return e.ForbiddenError("Requires admin role", nil)
+	}
+	if h.hb == nil {
+		return e.JSON(http.StatusOK, map[string]any{
+			"err": "Heartbeat not configured. Set HEARTBEAT_URL environment variable.",
+		})
+	}
+	if err := h.hb.Send(); err != nil {
+		return e.JSON(http.StatusOK, map[string]any{"err": err.Error()})
+	}
+	return e.JSON(http.StatusOK, map[string]any{"err": false})
+}
+
 // containerRequestHandler handles both container logs and info requests
 func (h *Hub) containerRequestHandler(e *core.RequestEvent, fetchFunc func(*systems.System, string) (string, error), responseKey string) error {
 	systemID := e.Request.URL.Query().Get("system")

internal/hub/hub_test.go

@@ -362,6 +362,58 @@ func TestApiRoutesAuthentication(t *testing.T) {
 			ExpectedContent: []string{"test-system"},
 			TestAppFactory:  testAppFactory,
 		},
+		{
+			Name:            "GET /heartbeat-status - no auth should fail",
+			Method:          http.MethodGet,
+			URL:             "/api/beszel/heartbeat-status",
+			ExpectedStatus:  401,
+			ExpectedContent: []string{"requires valid"},
+			TestAppFactory:  testAppFactory,
+		},
+		{
+			Name:   "GET /heartbeat-status - with user auth should fail",
+			Method: http.MethodGet,
+			URL:    "/api/beszel/heartbeat-status",
+			Headers: map[string]string{
+				"Authorization": userToken,
+			},
+			ExpectedStatus:  403,
+			ExpectedContent: []string{"Requires admin role"},
+			TestAppFactory:  testAppFactory,
+		},
+		{
+			Name:   "GET /heartbeat-status - with admin auth should succeed",
+			Method: http.MethodGet,
+			URL:    "/api/beszel/heartbeat-status",
+			Headers: map[string]string{
+				"Authorization": adminUserToken,
+			},
+			ExpectedStatus:  200,
+			ExpectedContent: []string{`"enabled":false`},
+			TestAppFactory:  testAppFactory,
+		},
+		{
+			Name:   "POST /test-heartbeat - with user auth should fail",
+			Method: http.MethodPost,
+			URL:    "/api/beszel/test-heartbeat",
+			Headers: map[string]string{
+				"Authorization": userToken,
+			},
+			ExpectedStatus:  403,
+			ExpectedContent: []string{"Requires admin role"},
+			TestAppFactory:  testAppFactory,
+		},
+		{
+			Name:   "POST /test-heartbeat - with admin auth should report disabled state",
+			Method: http.MethodPost,
+			URL:    "/api/beszel/test-heartbeat",
+			Headers: map[string]string{
+				"Authorization": adminUserToken,
+			},
+			ExpectedStatus:  200,
+			ExpectedContent: []string{"Heartbeat not configured"},
+			TestAppFactory:  testAppFactory,
+		},
 		{
 			Name:            "GET /universal-token - no auth should fail",
 			Method:          http.MethodGet,

internal/site/src/components/containers-table/containers-table.tsx

@@ -54,36 +54,34 @@ export default function ContainersTable({ systemId }: { systemId?: string }) {
 					fields: "id,name,image,cpu,memory,net,health,status,system,updated",
 					filter: systemId ? pb.filter("system={:system}", { system: systemId }) : undefined,
 				})
-				.then(
-					({ items }) => {
-						if (items.length === 0) {
-							setData((curItems) => {
-								if (systemId) {
-									return curItems?.filter((item) => item.system !== systemId) ?? []
-								}
-								return []
-							})
-							return
-						}
+				.then(({ items }) => {
+					if (items.length === 0) {
 						setData((curItems) => {
-							const lastUpdated = Math.max(items[0].updated, items.at(-1)?.updated ?? 0)
-							const containerIds = new Set()
-							const newItems = []
-							for (const item of items) {
-								if (Math.abs(lastUpdated - item.updated) < 70_000) {
-									containerIds.add(item.id)
-									newItems.push(item)
-								}
+							if (systemId) {
+								return curItems?.filter((item) => item.system !== systemId) ?? []
 							}
-							for (const item of curItems ?? []) {
-								if (!containerIds.has(item.id) && lastUpdated - item.updated < 70_000) {
-									newItems.push(item)
-								}
-							}
-							return newItems
+							return []
 						})
+						return
 					}
-				)
+					setData((curItems) => {
+						const lastUpdated = Math.max(items[0].updated, items.at(-1)?.updated ?? 0)
+						const containerIds = new Set()
+						const newItems = []
+						for (const item of items) {
+							if (Math.abs(lastUpdated - item.updated) < 70_000) {
+								containerIds.add(item.id)
+								newItems.push(item)
+							}
+						}
+						for (const item of curItems ?? []) {
+							if (!containerIds.has(item.id) && lastUpdated - item.updated < 70_000) {
+								newItems.push(item)
+							}
+						}
+						return newItems
+					})
+				})
 		}
 
 		// initial load
@@ -285,7 +283,7 @@ async function getInfoHtml(container: ContainerRecord): Promise<string> {
 		])
 		try {
 			info = JSON.stringify(JSON.parse(info), null, 2)
-		} catch (_) { }
+		} catch (_) {}
 		return info ? highlighter.codeToHtml(info, { lang: "json", theme: syntaxTheme }) : t`No results.`
 	} catch (error) {
 		console.error(error)
@@ -342,12 +340,12 @@ function ContainerSheet({
 		setLogsDisplay("")
 		setInfoDisplay("")
 		if (!container) return
-			; (async () => {
-				const [logsHtml, infoHtml] = await Promise.all([getLogsHtml(container), getInfoHtml(container)])
-				setLogsDisplay(logsHtml)
-				setInfoDisplay(infoHtml)
-				setTimeout(scrollLogsToBottom, 20)
-			})()
+		;(async () => {
+			const [logsHtml, infoHtml] = await Promise.all([getLogsHtml(container), getInfoHtml(container)])
+			setLogsDisplay(logsHtml)
+			setInfoDisplay(infoHtml)
+			setTimeout(scrollLogsToBottom, 20)
+		})()
 	}, [container])
 
 	return (
@@ -473,7 +471,7 @@ const ContainerTableRow = memo(function ContainerTableRow({
 			{row.getVisibleCells().map((cell) => (
 				<TableCell
 					key={cell.id}
-					className="py-0"
+					className="py-0 ps-4.5"
 					style={{
 						height: virtualRow.size,
 					}}

internal/site/src/components/routes/settings/heartbeat.tsx

@@ -0,0 +1,205 @@
+import { t } from "@lingui/core/macro"
+import { Trans } from "@lingui/react/macro"
+import { redirectPage } from "@nanostores/router"
+import clsx from "clsx"
+import { LoaderCircleIcon, SendIcon } from "lucide-react"
+import { useEffect, useState } from "react"
+import { $router } from "@/components/router"
+import { Badge } from "@/components/ui/badge"
+import { Button } from "@/components/ui/button"
+import { Separator } from "@/components/ui/separator"
+import { toast } from "@/components/ui/use-toast"
+import { isAdmin, pb } from "@/lib/api"
+
+interface HeartbeatStatus {
+	enabled: boolean
+	url?: string
+	interval?: number
+	method?: string
+	msg?: string
+}
+
+export default function HeartbeatSettings() {
+	const [status, setStatus] = useState<HeartbeatStatus | null>(null)
+	const [isLoading, setIsLoading] = useState(true)
+	const [isTesting, setIsTesting] = useState(false)
+
+	if (!isAdmin()) {
+		redirectPage($router, "settings", { name: "general" })
+	}
+
+	useEffect(() => {
+		fetchStatus()
+	}, [])
+
+	async function fetchStatus() {
+		try {
+			setIsLoading(true)
+			const res = await pb.send<HeartbeatStatus>("/api/beszel/heartbeat-status", {})
+			setStatus(res)
+		} catch (error: any) {
+			toast({
+				title: t`Error`,
+				description: error.message,
+				variant: "destructive",
+			})
+		} finally {
+			setIsLoading(false)
+		}
+	}
+
+	async function sendTestHeartbeat() {
+		setIsTesting(true)
+		try {
+			const res = await pb.send<{ err: string | false }>("/api/beszel/test-heartbeat", {
+				method: "POST",
+			})
+			if ("err" in res && !res.err) {
+				toast({
+					title: t`Heartbeat sent successfully`,
+					description: t`Check your monitoring service`,
+				})
+			} else {
+				toast({
+					title: t`Error`,
+					description: (res.err as string) ?? t`Failed to send heartbeat`,
+					variant: "destructive",
+				})
+			}
+		} catch (error: any) {
+			toast({
+				title: t`Error`,
+				description: error.message,
+				variant: "destructive",
+			})
+		} finally {
+			setIsTesting(false)
+		}
+	}
+
+	const TestIcon = isTesting ? LoaderCircleIcon : SendIcon
+
+	return (
+		<div>
+			<div>
+				<h3 className="text-xl font-medium mb-2">
+					<Trans>Heartbeat Monitoring</Trans>
+				</h3>
+				<p className="text-sm text-muted-foreground leading-relaxed">
+					<Trans>
+						Send periodic outbound pings to an external monitoring service so you can monitor Beszel without exposing it
+						to the internet.
+					</Trans>
+				</p>
+			</div>
+			<Separator className="my-4" />
+
+			{isLoading ? (
+				<div className="flex items-center gap-2 text-muted-foreground py-4">
+					<LoaderCircleIcon className="h-4 w-4 animate-spin" />
+					<Trans>Loading heartbeat status...</Trans>
+				</div>
+			) : status?.enabled ? (
+				<div className="space-y-5">
+					<div className="flex items-center gap-2">
+						<Badge variant="success">
+							<Trans>Active</Trans>
+						</Badge>
+					</div>
+					<div className="grid gap-4 sm:grid-cols-2">
+						<ConfigItem label={t`Endpoint URL`} value={status.url ?? ""} mono />
+						<ConfigItem label={t`Interval`} value={`${status.interval}s`} />
+						<ConfigItem label={t`HTTP Method`} value={status.method ?? "POST"} />
+					</div>
+
+					<Separator />
+
+					<div>
+						<h4 className="text-base font-medium mb-1">
+							<Trans>Test heartbeat</Trans>
+						</h4>
+						<p className="text-sm text-muted-foreground leading-relaxed mb-3">
+							<Trans>Send a single heartbeat ping to verify your endpoint is working.</Trans>
+						</p>
+						<Button
+							type="button"
+							variant="outline"
+							className="flex items-center gap-1.5"
+							onClick={sendTestHeartbeat}
+							disabled={isTesting}
+						>
+							<TestIcon className={clsx("h-4 w-4", isTesting && "animate-spin")} />
+							<Trans>Send test heartbeat</Trans>
+						</Button>
+					</div>
+
+					<Separator />
+
+					<div>
+						<h4 className="text-base font-medium mb-2">
+							<Trans>Payload format</Trans>
+						</h4>
+						<p className="text-sm text-muted-foreground leading-relaxed mb-2">
+							<Trans>
+								When using POST, each heartbeat includes a JSON payload with system status summary, list of down
+								systems, and triggered alerts.
+							</Trans>
+						</p>
+						<p className="text-sm text-muted-foreground leading-relaxed">
+							<Trans>
+								The overall status is <code className="bg-muted rounded-sm px-1 text-primary">ok</code> when all systems
+								are up, <code className="bg-muted rounded-sm px-1 text-primary">warn</code> when alerts are triggered,
+								and <code className="bg-muted rounded-sm px-1 text-primary">error</code> when any system is down.
+							</Trans>
+						</p>
+					</div>
+				</div>
+			) : (
+				<div className="grid gap-4">
+					<div>
+						<p className="text-sm text-muted-foreground leading-relaxed mb-3">
+							<Trans>Set the following environment variables on your Beszel hub to enable heartbeat monitoring:</Trans>
+						</p>
+						<div className="grid gap-2.5">
+							<EnvVarItem
+								name="HEARTBEAT_URL"
+								description={t`Endpoint URL to ping (required)`}
+								example="https://uptime.betterstack.com/api/v1/heartbeat/xxxx"
+							/>
+							<EnvVarItem name="HEARTBEAT_INTERVAL" description={t`Seconds between pings (default: 60)`} example="60" />
+							<EnvVarItem
+								name="HEARTBEAT_METHOD"
+								description={t`HTTP method: POST, GET, or HEAD (default: POST)`}
+								example="POST"
+							/>
+						</div>
+					</div>
+					<p className="text-sm text-muted-foreground leading-relaxed">
+						<Trans>After setting the environment variables, restart your Beszel hub for changes to take effect.</Trans>
+					</p>
+				</div>
+			)}
+		</div>
+	)
+}
+
+function ConfigItem({ label, value, mono }: { label: string; value: string; mono?: boolean }) {
+	return (
+		<div>
+			<p className="text-sm font-medium mb-0.5">{label}</p>
+			<p className={clsx("text-sm text-muted-foreground break-all", mono && "font-mono")}>{value}</p>
+		</div>
+	)
+}
+
+function EnvVarItem({ name, description, example }: { name: string; description: string; example: string }) {
+	return (
+		<div className="bg-muted/50 rounded-md px-3 py-2 grid gap-1.5">
+			<code className="text-sm font-mono text-primary font-medium leading-tight">{name}</code>
+			<p className="text-sm text-muted-foreground">{description}</p>
+			<p className="text-xs text-muted-foreground">
+				<Trans>Example:</Trans> <code className="font-mono">{example}</code>
+			</p>
+		</div>
+	)
+}

internal/site/src/components/routes/settings/layout.tsx

@@ -2,7 +2,14 @@ import { t } from "@lingui/core/macro"
 import { Trans, useLingui } from "@lingui/react/macro"
 import { useStore } from "@nanostores/react"
 import { getPagePath, redirectPage } from "@nanostores/router"
-import { AlertOctagonIcon, BellIcon, FileSlidersIcon, FingerprintIcon, SettingsIcon } from "lucide-react"
+import {
+	AlertOctagonIcon,
+	BellIcon,
+	FileSlidersIcon,
+	FingerprintIcon,
+	HeartPulseIcon,
+	SettingsIcon,
+} from "lucide-react"
 import { lazy, useEffect } from "react"
 import { $router } from "@/components/router.tsx"
 import { Card, CardContent, CardDescription, CardHeader, CardTitle } from "@/components/ui/card.tsx"
@@ -18,12 +25,14 @@ const notificationsSettingsImport = () => import("./notifications.tsx")
 const configYamlSettingsImport = () => import("./config-yaml.tsx")
 const fingerprintsSettingsImport = () => import("./tokens-fingerprints.tsx")
 const alertsHistoryDataTableSettingsImport = () => import("./alerts-history-data-table.tsx")
+const heartbeatSettingsImport = () => import("./heartbeat.tsx")
 
 const GeneralSettings = lazy(generalSettingsImport)
 const NotificationsSettings = lazy(notificationsSettingsImport)
 const ConfigYamlSettings = lazy(configYamlSettingsImport)
 const FingerprintsSettings = lazy(fingerprintsSettingsImport)
 const AlertsHistoryDataTableSettings = lazy(alertsHistoryDataTableSettingsImport)
+const HeartbeatSettings = lazy(heartbeatSettingsImport)
 
 export async function saveSettings(newSettings: Partial<UserSettings>) {
 	try {
@@ -88,6 +97,13 @@ export default function SettingsLayout() {
 			admin: true,
 			preload: configYamlSettingsImport,
 		},
+		{
+			title: t`Heartbeat`,
+			href: getPagePath($router, "settings", { name: "heartbeat" }),
+			icon: HeartPulseIcon,
+			admin: true,
+			preload: heartbeatSettingsImport,
+		},
 	]
 
 	const page = useStore($router)
@@ -141,5 +157,7 @@ function SettingsContent({ name }: { name: string }) {
 			return <FingerprintsSettings />
 		case "alert-history":
 			return <AlertsHistoryDataTableSettings />
+		case "heartbeat":
+			return <HeartbeatSettings />
 	}
 }

internal/site/src/components/routes/system/info-bar.tsx

@@ -19,7 +19,7 @@ import { FreeBsdIcon, TuxIcon, WebSocketIcon, WindowsIcon } from "@/components/u
 import { Separator } from "@/components/ui/separator"
 import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip"
 import { ConnectionType, connectionTypeLabels, Os, SystemStatus } from "@/lib/enums"
-import { cn, formatBytes, getHostDisplayValue, secondsToString, toFixedFloat } from "@/lib/utils"
+import { cn, formatBytes, getHostDisplayValue, secondsToUptimeString, toFixedFloat } from "@/lib/utils"
 import type { ChartData, SystemDetailsRecord, SystemRecord } from "@/types"
 
 export default function InfoBar({
@@ -77,14 +77,6 @@ export default function InfoBar({
 			},
 		}
 
-		let uptime: string
-		if (system.info.u < 3600) {
-			uptime = secondsToString(system.info.u, "minute")
-		} else if (system.info.u < 360000) {
-			uptime = secondsToString(system.info.u, "hour")
-		} else {
-			uptime = secondsToString(system.info.u, "day")
-		}
 		const info = [
 			{ value: getHostDisplayValue(system), Icon: GlobeIcon },
 			{
@@ -94,7 +86,7 @@ export default function InfoBar({
 				// hide if hostname is same as host or name
 				hide: hostname === system.host || hostname === system.name,
 			},
-			{ value: uptime, Icon: ClockArrowUp, label: t`Uptime`, hide: !system.info.u },
+			{ value: secondsToUptimeString(system.info.u), Icon: ClockArrowUp, label: t`Uptime`, hide: !system.info.u },
 			osInfo[os],
 			{
 				value: cpuModel,

internal/site/src/components/routes/system/smart-table.tsx

@@ -174,8 +174,8 @@ export const columns: ColumnDef<SmartDeviceRecord>[] = [
 			<HeaderButton column={column} name={t({ message: "Power On", comment: "Power On Time" })} Icon={Clock} />
 		),
 		cell: ({ getValue }) => {
-			const hours = (getValue() ?? 0) as number
-			if (!hours && hours !== 0) {
+			const hours = getValue() as number | undefined
+			if (hours == null) {
 				return <div className="text-sm text-muted-foreground ms-1.5">N/A</div>
 			}
 			const seconds = hours * 3600
@@ -195,7 +195,7 @@ export const columns: ColumnDef<SmartDeviceRecord>[] = [
 		),
 		cell: ({ getValue }) => {
 			const cycles = getValue() as number | undefined
-			if (!cycles && cycles !== 0) {
+			if (cycles == null) {
 				return <div className="text-muted-foreground ms-1.5">N/A</div>
 			}
 			return <span className="ms-1.5">{cycles.toLocaleString()}</span>
@@ -206,7 +206,11 @@ export const columns: ColumnDef<SmartDeviceRecord>[] = [
 		invertSorting: true,
 		header: ({ column }) => <HeaderButton column={column} name={t`Temp`} Icon={ThermometerIcon} />,
 		cell: ({ getValue }) => {
-			const { value, unit } = formatTemperature(getValue() as number)
+			const temp = getValue() as number | null | undefined
+			if (!temp) {
+				return <div className="text-muted-foreground ms-1.5">N/A</div>
+			}
+			const { value, unit } = formatTemperature(temp)
 			return <span className="ms-1.5">{`${value} ${unit}`}</span>
 		},
 	},
@@ -304,41 +308,41 @@ export default function DisksTable({ systemId }: { systemId?: string }) {
 			? { fields: SMART_DEVICE_FIELDS, filter: pb.filter("system = {:system}", { system: systemId }) }
 			: { fields: SMART_DEVICE_FIELDS }
 
-			; (async () => {
-				try {
-					unsubscribe = await pb.collection("smart_devices").subscribe(
-						"*",
-						(event) => {
-							const record = event.record as SmartDeviceRecord
-							setSmartDevices((currentDevices) => {
-								const devices = currentDevices ?? []
-								const matchesSystemScope = !systemId || record.system === systemId
+		;(async () => {
+			try {
+				unsubscribe = await pb.collection("smart_devices").subscribe(
+					"*",
+					(event) => {
+						const record = event.record as SmartDeviceRecord
+						setSmartDevices((currentDevices) => {
+							const devices = currentDevices ?? []
+							const matchesSystemScope = !systemId || record.system === systemId
 
-								if (event.action === "delete") {
-									return devices.filter((device) => device.id !== record.id)
-								}
+							if (event.action === "delete") {
+								return devices.filter((device) => device.id !== record.id)
+							}
 
-								if (!matchesSystemScope) {
-									// Record moved out of scope; ensure it disappears locally.
-									return devices.filter((device) => device.id !== record.id)
-								}
+							if (!matchesSystemScope) {
+								// Record moved out of scope; ensure it disappears locally.
+								return devices.filter((device) => device.id !== record.id)
+							}
 
-								const existingIndex = devices.findIndex((device) => device.id === record.id)
-								if (existingIndex === -1) {
-									return [record, ...devices]
-								}
+							const existingIndex = devices.findIndex((device) => device.id === record.id)
+							if (existingIndex === -1) {
+								return [record, ...devices]
+							}
 
-								const next = [...devices]
-								next[existingIndex] = record
-								return next
-							})
-						},
-						pbOptions
-					)
-				} catch (error) {
-					console.error("Failed to subscribe to SMART device updates:", error)
-				}
-			})()
+							const next = [...devices]
+							next[existingIndex] = record
+							return next
+						})
+					},
+					pbOptions
+				)
+			} catch (error) {
+				console.error("Failed to subscribe to SMART device updates:", error)
+			}
+		})()
 
 		return () => {
 			unsubscribe?.()

internal/site/src/components/systems-table/systems-table-columns.tsx

@@ -35,7 +35,7 @@ import {
 	formatTemperature,
 	getMeterState,
 	parseSemVer,
-	secondsToString,
+	secondsToUptimeString,
 } from "@/lib/utils"
 import { batteryStateTranslations } from "@/lib/i18n"
 import type { SystemRecord } from "@/types"
@@ -154,11 +154,7 @@ export function SystemsTableColumns(viewMode: "table" | "grid"): ColumnDef<Syste
 								{name}
 							</Link>
 						</span>
-						<Link
-							href={linkUrl}
-							className="inset-0 absolute size-full"
-							aria-label={name}
-						></Link>
+						<Link href={linkUrl} className="inset-0 absolute size-full" aria-label={name}></Link>
 					</>
 				)
 			},
@@ -382,20 +378,13 @@ export function SystemsTableColumns(viewMode: "table" | "grid"): ColumnDef<Syste
 			size: 50,
 			Icon: ClockArrowUp,
 			header: sortableHeader,
+			hideSort: true,
 			cell(info) {
 				const uptime = info.getValue() as number
 				if (!uptime) {
 					return null
 				}
-				let formatted: string
-				if (uptime < 3600) {
-					formatted = secondsToString(uptime, "minute")
-				} else if (uptime < 360000) {
-					formatted = secondsToString(uptime, "hour")
-				} else {
-					formatted = secondsToString(uptime, "day")
-				}
-				return <span className="tabular-nums whitespace-nowrap">{formatted}</span>
+				return <span className="tabular-nums whitespace-nowrap">{secondsToUptimeString(uptime)}</span>
 			},
 		},
 		{
@@ -479,9 +468,9 @@ function TableCellWithMeter(info: CellContext<SystemRecord, unknown>) {
 	const meterClass = cn(
 		"h-full",
 		(info.row.original.status !== SystemStatus.Up && STATUS_COLORS.paused) ||
-		(threshold === MeterState.Good && STATUS_COLORS.up) ||
-		(threshold === MeterState.Warn && STATUS_COLORS.pending) ||
-		STATUS_COLORS.down
+			(threshold === MeterState.Good && STATUS_COLORS.up) ||
+			(threshold === MeterState.Warn && STATUS_COLORS.pending) ||
+			STATUS_COLORS.down
 	)
 	return (
 		<div className="flex gap-2 items-center tabular-nums tracking-tight w-full">
@@ -593,7 +582,7 @@ export function IndicatorDot({ system, className }: { system: SystemRecord; clas
 	return (
 		<span
 			className={cn("shrink-0 size-2 rounded-full", className)}
-		// style={{ marginBottom: "-1px" }}
+			// style={{ marginBottom: "-1px" }}
 		/>
 	)
 }

internal/site/src/components/systems-table/systems-table.tsx

@@ -434,7 +434,7 @@ const SystemTableRow = memo(
 								width: cell.column.getSize(),
 								height: virtualRow.size,
 							}}
-							className="py-0"
+							className="py-0 ps-4.5"
 						>
 							{flexRender(cell.column.columnDef.cell, cell.getContext())}
 						</TableCell>

internal/site/src/lib/utils.ts

@@ -465,4 +465,15 @@ export function secondsToString(seconds: number, unit: "hour" | "minute" | "day"
 		case "day":
 			return plural(count, { one: `${countString} day`, other: `${countString} days` })
 	}
+}
+
+/** Format seconds to uptime string - "X minutes", "X hours", "X days" */
+export function secondsToUptimeString(seconds: number): string {
+	if (seconds < 3600) {
+		return secondsToString(seconds, "minute")
+	} else if (seconds < 360000) {
+		return secondsToString(seconds, "hour")
+	} else {
+		return secondsToString(seconds, "day")
+	}
 }

readme.md

@@ -51,7 +51,7 @@ The [quick start guide](https://beszel.dev/guide/getting-started) and other docu
 - **GPU usage / power draw** - Nvidia, AMD, and Intel.
 - **Battery** - Host system battery charge.
 - **Containers** - Status and metrics of all running Docker / Podman containers.
-- **S.M.A.R.T.** - Host system disk health.
+- **S.M.A.R.T.** - Host system disk health (includes eMMC wear/EOL via Linux sysfs when available).
 
 ## Help and discussion
 

supplemental/CHANGELOG.md

@@ -1,3 +1,9 @@
+## Unreleased
+
+- Add outbound heartbeat monitoring to external services (BetterStack, Uptime Kuma, Healthchecks.io, etc.) with system status summary payload. Configured via `BESZEL_HUB_HEARTBEAT_URL`, `BESZEL_HUB_HEARTBEAT_INTERVAL`, and `BESZEL_HUB_HEARTBEAT_METHOD` environment variables.
+
+- Add Heartbeat settings page to the admin UI with status display, configuration reference, and test button.
+
 ## 0.18.3
 
 - Add experimental sysfs AMD GPU collector. (#737, #1569)