Update power benchmark GPU reset flow

audit/internal/platform/benchmark.go

@@ -37,6 +37,8 @@ type benchmarkGPUInfo struct {
 	VBIOS                string
 	PowerLimitW          float64
 	DefaultPowerLimitW   float64
+	MinPowerLimitW       float64
+	MaxPowerLimitW       float64
 	MaxGraphicsClockMHz  float64
 	MaxMemoryClockMHz    float64
 	BaseGraphicsClockMHz float64
@@ -95,6 +97,8 @@ var (
 	benchmarkReadyPattern      = regexp.MustCompile(`^([a-z0-9_]+)\[(\d+)\]=READY dim=(\d+)x(\d+)x(\d+)\b`)
 	benchmarkSkippedPattern    = regexp.MustCompile(`^([a-z0-9_]+)(?:\[\d+\])?=SKIPPED (.+)$`)
 	benchmarkIterationsPattern = regexp.MustCompile(`^([a-z0-9_]+)_iterations=(\d+)$`)
+	benchmarkGeteuid           = os.Geteuid
+	benchmarkSleep             = time.Sleep
 )
 
 // benchmarkPrecisionPhases lists the precision categories run as individual
@@ -220,8 +224,6 @@ func benchmarkCalibrationThrottleReason(before, after BenchmarkThrottleCounters)
 		return "hw_thermal"
 	case diff.SWThermalSlowdownUS > 0:
 		return "sw_thermal"
-	case diff.HWPowerBrakeSlowdownUS > 0:
-		return "hw_power_brake"
 	default:
 		return ""
 	}
@@ -240,6 +242,39 @@ func setBenchmarkPowerLimit(ctx context.Context, verboseLog string, gpuIndex, po
 	return nil
 }
 
+func resetBenchmarkGPUs(ctx context.Context, verboseLog string, gpuIndices []int, logFunc func(string)) []int {
+	if len(gpuIndices) == 0 {
+		return nil
+	}
+	if benchmarkGeteuid() != 0 {
+		if logFunc != nil {
+			logFunc("power benchmark pre-flight: root privileges unavailable, GPU reset skipped")
+		}
+		return append([]int(nil), gpuIndices...)
+	}
+	if killed := KillTestWorkers(); len(killed) > 0 && logFunc != nil {
+		for _, p := range killed {
+			logFunc(fmt.Sprintf("power benchmark pre-flight: killed stale worker pid=%d name=%s", p.PID, p.Name))
+		}
+	}
+	var failed []int
+	for _, idx := range gpuIndices {
+		name := fmt.Sprintf("power-preflight-gpu-%d-reset.log", idx)
+		if _, err := runSATCommandCtx(ctx, verboseLog, name, []string{"nvidia-smi", "-i", strconv.Itoa(idx), "-r"}, nil, logFunc); err != nil {
+			failed = append(failed, idx)
+			if logFunc != nil {
+				logFunc(fmt.Sprintf("power benchmark pre-flight: GPU %d reset failed: %v", idx, err))
+			}
+			continue
+		}
+		if logFunc != nil {
+			logFunc(fmt.Sprintf("power benchmark pre-flight: GPU %d reset completed", idx))
+		}
+		benchmarkSleep(time.Second)
+	}
+	return failed
+}
+
 func benchmarkPowerEngine() string {
 	switch strings.TrimSpace(strings.ToLower(os.Getenv("BEE_BENCH_POWER_ENGINE"))) {
 	case BenchmarkPowerEngineTargetedPower:
@@ -351,9 +386,9 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		result.Warnings = append(result.Warnings, "gpu inventory query failed: "+infoErr.Error())
 		result.Normalization.Status = "partial"
 	}
-	// Enrich with max clocks from verbose output — covers GPUs where
-	// clocks.max.* CSV fields are unsupported (e.g. Blackwell / driver 98.x).
-	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQOut)
+	// Enrich with verbose nvidia-smi data — covers GPUs where some CSV fields
+	// are unsupported (e.g. clocks.max.* on Blackwell / driver 98.x).
+	enrichGPUInfoWithNvidiaSMIQ(infoByIndex, nvsmiQOut)
 
 	activeApps, err := queryActiveComputeApps(selected)
 	if err == nil && len(activeApps) > 0 {
@@ -737,8 +772,8 @@ func resolveBenchmarkProfile(profile string) benchmarkProfileSpec {
 // (attribute.multiprocessor_count, power.default_limit) are not supported on
 // all driver versions, so we fall back to the base set if the full query fails.
 // The minimal fallback omits clock fields entirely — clocks.max.* returns
-// exit status 2 on some GPU generations (e.g. Blackwell); max clocks are
-// then recovered from nvidia-smi -q via enrichGPUInfoWithMaxClocks.
+// exit status 2 on some GPU generations (e.g. Blackwell); missing data is
+// then recovered from nvidia-smi -q.
 var benchmarkGPUInfoQueries = []struct {
 	fields   string
 	extended bool // whether this query includes optional extended fields
@@ -758,12 +793,9 @@ var benchmarkGPUInfoQueries = []struct {
 	},
 }
 
-// enrichGPUInfoWithMaxClocks fills MaxGraphicsClockMHz / MaxMemoryClockMHz for
-// any GPU in infoByIndex where those values are still zero.  It parses the
-// "Max Clocks" section of nvidia-smi -q output (already available as nvsmiQ).
-// This is the fallback for GPUs (e.g. Blackwell) where clocks.max.* CSV fields
-// return exit status 2 but the verbose query works fine.
-func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []byte) {
+// enrichGPUInfoWithNvidiaSMIQ fills benchmark GPU metadata from nvidia-smi -q
+// for fields that may be missing from --query-gpu on some driver versions.
+func enrichGPUInfoWithNvidiaSMIQ(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []byte) {
 	if len(infoByIndex) == 0 || len(nvsmiQ) == 0 {
 		return
 	}
@@ -784,6 +816,8 @@ func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []b
 	maxMemRe := regexp.MustCompile(`(?i)Max Clocks[\s\S]*?Memory\s*:\s*(\d+)\s*MHz`)
 	defaultPwrRe := regexp.MustCompile(`(?i)Default Power Limit\s*:\s*([0-9.]+)\s*W`)
 	currentPwrRe := regexp.MustCompile(`(?i)Current Power Limit\s*:\s*([0-9.]+)\s*W`)
+	minPwrRe := regexp.MustCompile(`(?i)Min Power Limit\s*:\s*([0-9.]+)\s*W`)
+	maxPwrRe := regexp.MustCompile(`(?i)Max Power Limit\s*:\s*([0-9.]+)\s*W`)
 	smCountRe := regexp.MustCompile(`(?i)Multiprocessor Count\s*:\s*(\d+)`)
 	shutdownTempRe := regexp.MustCompile(`(?i)GPU Shutdown Temp\s*:\s*(\d+)\s*C`)
 	slowdownTempRe := regexp.MustCompile(`(?i)GPU Slowdown Temp\s*:\s*(\d+)\s*C`)
@@ -843,6 +877,20 @@ func enrichGPUInfoWithMaxClocks(infoByIndex map[int]benchmarkGPUInfo, nvsmiQ []b
 				}
 			}
 		}
+		if info.MinPowerLimitW == 0 {
+			if m := minPwrRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil && v > 0 {
+					info.MinPowerLimitW = v
+				}
+			}
+		}
+		if info.MaxPowerLimitW == 0 {
+			if m := maxPwrRe.FindSubmatch(section); m != nil {
+				if v, err := strconv.ParseFloat(string(m[1]), 64); err == nil && v > 0 {
+					info.MaxPowerLimitW = v
+				}
+			}
+		}
 		if info.MultiprocessorCount == 0 {
 			if m := smCountRe.FindSubmatch(section); m != nil {
 				if v, err := strconv.Atoi(string(m[1])); err == nil && v > 0 {
@@ -3043,7 +3091,6 @@ func runBenchmarkPowerCalibration(
 	if calibDurationSec <= 0 {
 		calibDurationSec = 120
 	}
-	const maxDerateW = 150
 	// calibSearchTolerance is the binary-search convergence threshold in watts.
 	// When hi-lo ≤ this, the highest verified-stable limit (lo) is used.
 	const calibSearchTolerance = 10
@@ -3090,8 +3137,9 @@ func runBenchmarkPowerCalibration(
 		originalLimitW int
 		appliedLimitW  int
 		minLimitW      int
-		lo             int // highest verified-stable limit (assumed: minLimitW)
+		lo             int // highest verified-stable limit
 		hi             int // lowest verified-unstable limit (exclusive sentinel above start)
+		loVerified     bool
 		calib          benchmarkPowerCalibrationResult
 		converged      bool
 	}
@@ -3113,23 +3161,17 @@ func runBenchmarkPowerCalibration(
 		if defaultLimitW <= 0 {
 			defaultLimitW = originalLimitW
 		}
-		appliedLimitW := originalLimitW
+		appliedLimitW := initialBenchmarkCalibrationLimitW(info)
 		if appliedLimitW <= 0 {
 			appliedLimitW = defaultLimitW
 		}
-		minLimitW := appliedLimitW
-		switch {
-		case defaultLimitW > 0:
-			minLimitW = defaultLimitW - maxDerateW
-			floorByRatio := int(math.Round(float64(defaultLimitW) * 0.70))
-			if minLimitW < floorByRatio {
-				minLimitW = floorByRatio
-			}
-		case appliedLimitW > 0:
-			minLimitW = appliedLimitW - maxDerateW
+		minLimitW := int(math.Round(info.MinPowerLimitW))
+		if minLimitW <= 0 {
+			minLimitW = appliedLimitW
 		}
-		if minLimitW < calibSearchTolerance {
-			minLimitW = calibSearchTolerance
+		maxLimitW := int(math.Round(info.MaxPowerLimitW))
+		if maxLimitW > 0 && appliedLimitW > maxLimitW {
+			appliedLimitW = maxLimitW
 		}
 		s := &gpuCalibState{
 			idx:            idx,
@@ -3141,11 +3183,24 @@ func runBenchmarkPowerCalibration(
 			hi:             appliedLimitW + 1, // not yet tested, not yet confirmed unstable
 			calib:          benchmarkPowerCalibrationResult{AppliedPowerLimitW: float64(appliedLimitW)},
 		}
+		if minLimitW > 0 && appliedLimitW > 0 && minLimitW >= appliedLimitW {
+			s.appliedLimitW = minLimitW
+			s.hi = minLimitW + 1
+		}
+		if info.MinPowerLimitW <= 0 {
+			s.calib.Notes = append(s.calib.Notes, "minimum power limit was not reported by nvidia-smi; calibration can only validate the current/default power limit")
+		}
 		if seedLimits != nil {
 			if seedW, ok := seedLimits[idx]; ok && seedW > 0 {
 				// A previously validated limit is only a starting point. Re-run
 				// targeted_power under the current multi-GPU thermal load and derate
 				// again if this step shows new throttling.
+				if seedW < s.minLimitW {
+					seedW = s.minLimitW
+				}
+				if maxLimitW > 0 && seedW > maxLimitW {
+					seedW = maxLimitW
+				}
 				if canDerate {
 					_ = setBenchmarkPowerLimit(ctx, verboseLog, idx, seedW)
 				}
@@ -3333,6 +3388,7 @@ calibDone:
 				s.calib.AppliedPowerLimitW = float64(s.appliedLimitW)
 				logFunc(fmt.Sprintf("power calibration: GPU %d stable at %d W, p95=%.0f W p95_temp=%.1f C (%d samples)", s.idx, s.appliedLimitW, summary.P95PowerW, summary.P95TempC, summary.Samples))
 				s.lo = s.appliedLimitW
+				s.loVerified = true
 				if canDerate && s.hi-s.lo > calibSearchTolerance {
 					next := roundTo5W((s.lo + s.hi) / 2)
 					if next > s.lo && next < s.hi {
@@ -3371,7 +3427,23 @@ calibDone:
 			s.hi = s.appliedLimitW
 
 			if s.hi-s.lo <= calibSearchTolerance {
-				if s.lo > s.minLimitW {
+				if !s.loVerified && s.minLimitW > 0 && s.appliedLimitW != s.minLimitW {
+					if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, s.minLimitW); err != nil {
+						s.calib.Notes = append(s.calib.Notes, "failed to set power limit: "+err.Error())
+						logFunc(fmt.Sprintf("power calibration: GPU %d failed to set minimum power limit %d W: %v", s.idx, s.minLimitW, err))
+						s.converged = true
+						continue
+					}
+					s.appliedLimitW = s.minLimitW
+					s.calib.AppliedPowerLimitW = float64(s.minLimitW)
+					s.calib.Derated = s.minLimitW < s.originalLimitW
+					s.info.PowerLimitW = float64(s.minLimitW)
+					infoByIndex[s.idx] = s.info
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: validating minimum settable limit %d W before concluding failure", s.minLimitW))
+					logFunc(fmt.Sprintf("power calibration: GPU %d binary search: validating minimum settable limit %d W", s.idx, s.minLimitW))
+					continue
+				}
+				if s.loVerified {
 					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search converged: using %d W (lo=%d hi=%d)", s.lo, s.lo, s.hi))
 					if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, s.lo); err == nil {
 						s.appliedLimitW = s.lo
@@ -3383,7 +3455,8 @@ calibDone:
 						s.calib.Completed = true
 					}
 				} else {
-					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable %s limit within %d W of the default", engineLabel, maxDerateW))
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable %s limit down to the minimum settable power limit %d W", engineLabel, s.minLimitW))
+					logFunc(fmt.Sprintf("power calibration: GPU %d no stable limit found down to minimum settable power limit %d W", s.idx, s.minLimitW))
 				}
 				s.calib.MetricRows = filterRowsByGPU(ar.rows, s.idx)
 				s.converged = true
@@ -3398,9 +3471,7 @@ calibDone:
 				next = (s.lo + s.hi) / 2
 			}
 			if next < s.minLimitW {
-				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable %s limit within %d W of the default", engineLabel, maxDerateW))
-				s.converged = true
-				continue
+				next = s.minLimitW
 			}
 			if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, next); err != nil {
 				s.calib.Notes = append(s.calib.Notes, "failed to set power limit: "+err.Error())
@@ -3439,6 +3510,24 @@ func roundTo5W(w int) int {
 	return ((w + 2) / 5) * 5
 }
 
+func initialBenchmarkCalibrationLimitW(info benchmarkGPUInfo) int {
+	defaultLimitW := int(math.Round(info.DefaultPowerLimitW))
+	currentLimitW := int(math.Round(info.PowerLimitW))
+	maxLimitW := int(math.Round(info.MaxPowerLimitW))
+
+	startW := defaultLimitW
+	if startW <= 0 {
+		startW = currentLimitW
+	}
+	if startW <= 0 {
+		startW = maxLimitW
+	}
+	if maxLimitW > 0 && startW > maxLimitW {
+		startW = maxLimitW
+	}
+	return startW
+}
+
 // meanFanRPM returns the average RPM across a set of fan readings.
 func meanFanRPM(fans []FanReading) float64 {
 	if len(fans) == 0 {
@@ -4096,14 +4185,6 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		return "", fmt.Errorf("mkdir %s: %w", runDir, err)
 	}
 	verboseLog := filepath.Join(runDir, "verbose.log")
-	infoByIndex, infoErr := queryBenchmarkGPUInfo(selected)
-	if infoErr != nil {
-		return "", infoErr
-	}
-	// Capture full nvidia-smi -q snapshot at the start of the run.
-	if out, err := runSATCommandCtx(ctx, verboseLog, "00-nvidia-smi-q.log", []string{"nvidia-smi", "-q"}, nil, nil); err == nil {
-		_ = os.WriteFile(filepath.Join(runDir, "00-nvidia-smi-q.log"), out, 0644)
-	}
 	hostname, _ := os.Hostname()
 	result := NvidiaPowerBenchResult{
 		BenchmarkVersion:   benchmarkVersion,
@@ -4114,6 +4195,14 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		SelectedGPUIndices: append([]int(nil), selected...),
 		OverallStatus:      "OK",
 	}
+	infoByIndex, infoErr := queryBenchmarkGPUInfo(selected)
+	if infoErr != nil {
+		return "", infoErr
+	}
+	// Capture full nvidia-smi -q snapshot at the start of the run.
+	if out, err := runSATCommandCtx(ctx, verboseLog, "00-nvidia-smi-q.log", []string{"nvidia-smi", "-q"}, nil, nil); err == nil {
+		_ = os.WriteFile(filepath.Join(runDir, "00-nvidia-smi-q.log"), out, 0644)
+	}
 	durationSec := powerBenchDurationSec(opts.Profile)
 
 	// Sample server idle power before any GPU load.
@@ -4139,6 +4228,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		singleDir := filepath.Join(runDir, fmt.Sprintf("single-%02d", idx))
 		_ = os.MkdirAll(singleDir, 0755)
 		singleInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
+		if failed := resetBenchmarkGPUs(ctx, verboseLog, []int{idx}, logFunc); len(failed) > 0 {
+			result.Findings = append(result.Findings,
+				fmt.Sprintf("GPU %d reset pre-flight did not complete before its first power test; throttle counters may contain stale state.", idx))
+		}
 		logFunc(fmt.Sprintf("power calibration: GPU %d single-card baseline", idx))
 		singlePowerStopCh := make(chan struct{})
 		singlePowerCh := startSelectedPowerSourceSampler(singlePowerStopCh, opts.ServerPowerSource, benchmarkPowerAutotuneSampleInterval)

audit/internal/platform/benchmark_test.go

@@ -1,8 +1,13 @@
 package platform
 
 import (
+	"context"
+	"os"
+	"os/exec"
+	"path/filepath"
 	"strings"
 	"testing"
+	"time"
 )
 
 func TestResolveBenchmarkProfile(t *testing.T) {
@@ -164,6 +169,93 @@ func TestBenchmarkPlannedPhaseStatus(t *testing.T) {
 	}
 }
 
+func TestBenchmarkCalibrationThrottleReasonIgnoresPowerReasons(t *testing.T) {
+	t.Parallel()
+
+	before := BenchmarkThrottleCounters{}
+	if got := benchmarkCalibrationThrottleReason(before, BenchmarkThrottleCounters{SWPowerCapUS: 1_000_000}); got != "" {
+		t.Fatalf("sw_power_cap should be ignored, got %q", got)
+	}
+	if got := benchmarkCalibrationThrottleReason(before, BenchmarkThrottleCounters{HWPowerBrakeSlowdownUS: 1_000_000}); got != "" {
+		t.Fatalf("hw_power_brake should be ignored, got %q", got)
+	}
+	if got := benchmarkCalibrationThrottleReason(before, BenchmarkThrottleCounters{HWThermalSlowdownUS: 1_000_000}); got != "hw_thermal" {
+		t.Fatalf("hw_thermal mismatch: got %q", got)
+	}
+	if got := benchmarkCalibrationThrottleReason(before, BenchmarkThrottleCounters{SWThermalSlowdownUS: 1_000_000}); got != "sw_thermal" {
+		t.Fatalf("sw_thermal mismatch: got %q", got)
+	}
+}
+
+func TestResetBenchmarkGPUsSkipsWithoutRoot(t *testing.T) {
+	t.Parallel()
+
+	oldGeteuid := benchmarkGeteuid
+	oldExec := satExecCommand
+	benchmarkGeteuid = func() int { return 1000 }
+	satExecCommand = func(name string, args ...string) *exec.Cmd {
+		t.Fatalf("unexpected command: %s %v", name, args)
+		return nil
+	}
+	t.Cleanup(func() {
+		benchmarkGeteuid = oldGeteuid
+		satExecCommand = oldExec
+	})
+
+	var logs []string
+	failed := resetBenchmarkGPUs(context.Background(), filepath.Join(t.TempDir(), "verbose.log"), []int{0, 2}, func(line string) {
+		logs = append(logs, line)
+	})
+	if got, want := strings.Join(logs, "\n"), "power benchmark pre-flight: root privileges unavailable, GPU reset skipped"; !strings.Contains(got, want) {
+		t.Fatalf("logs=%q want substring %q", got, want)
+	}
+	if len(failed) != 2 || failed[0] != 0 || failed[1] != 2 {
+		t.Fatalf("failed=%v want [0 2]", failed)
+	}
+}
+
+func TestResetBenchmarkGPUsResetsEachGPU(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	script := filepath.Join(dir, "nvidia-smi")
+	argsLog := filepath.Join(dir, "args.log")
+	if err := os.WriteFile(script, []byte("#!/bin/sh\nprintf '%s\\n' \"$*\" >> "+argsLog+"\nprintf 'ok\\n'\n"), 0755); err != nil {
+		t.Fatalf("write script: %v", err)
+	}
+
+	oldGeteuid := benchmarkGeteuid
+	oldSleep := benchmarkSleep
+	oldLookPath := satLookPath
+	benchmarkGeteuid = func() int { return 0 }
+	benchmarkSleep = func(time.Duration) {}
+	satLookPath = func(file string) (string, error) {
+		if file == "nvidia-smi" {
+			return script, nil
+		}
+		return exec.LookPath(file)
+	}
+	t.Cleanup(func() {
+		benchmarkGeteuid = oldGeteuid
+		benchmarkSleep = oldSleep
+		satLookPath = oldLookPath
+	})
+
+	failed := resetBenchmarkGPUs(context.Background(), filepath.Join(dir, "verbose.log"), []int{2, 5}, nil)
+	if len(failed) != 0 {
+		t.Fatalf("failed=%v want no failures", failed)
+	}
+	raw, err := os.ReadFile(argsLog)
+	if err != nil {
+		t.Fatalf("read args log: %v", err)
+	}
+	got := strings.Fields(string(raw))
+	want := []string{"-i", "2", "-r", "-i", "5", "-r"}
+	if strings.Join(got, " ") != strings.Join(want, " ") {
+		t.Fatalf("args=%v want %v", got, want)
+	}
+}
+
 func TestNormalizeNvidiaBenchmarkOptionsPreservesRunNCCLChoice(t *testing.T) {
 	t.Parallel()
 
@@ -179,6 +271,59 @@ func TestNormalizeNvidiaBenchmarkOptionsPreservesRunNCCLChoice(t *testing.T) {
 	}
 }
 
+func TestInitialBenchmarkCalibrationLimitW(t *testing.T) {
+	t.Parallel()
+
+	cases := []struct {
+		name string
+		info benchmarkGPUInfo
+		want int
+	}{
+		{
+			name: "prefers default tdp over current derated limit",
+			info: benchmarkGPUInfo{
+				PowerLimitW:        500,
+				DefaultPowerLimitW: 600,
+				MaxPowerLimitW:     600,
+			},
+			want: 600,
+		},
+		{
+			name: "caps default tdp to reported max limit",
+			info: benchmarkGPUInfo{
+				PowerLimitW:        500,
+				DefaultPowerLimitW: 700,
+				MaxPowerLimitW:     650,
+			},
+			want: 650,
+		},
+		{
+			name: "falls back to current limit when default missing",
+			info: benchmarkGPUInfo{
+				PowerLimitW:    525,
+				MaxPowerLimitW: 600,
+			},
+			want: 525,
+		},
+		{
+			name: "falls back to max limit when only that is known",
+			info: benchmarkGPUInfo{
+				MaxPowerLimitW: 575,
+			},
+			want: 575,
+		},
+	}
+
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			if got := initialBenchmarkCalibrationLimitW(tc.info); got != tc.want {
+				t.Fatalf("initialBenchmarkCalibrationLimitW(%+v)=%d want %d", tc.info, got, tc.want)
+			}
+		})
+	}
+}
+
 func TestParseBenchmarkBurnLog(t *testing.T) {
 	t.Parallel()
 
@@ -338,12 +483,16 @@ func TestScoreBenchmarkGPUIgnoresDisabledPrecisions(t *testing.T) {
 	}
 }
 
-func TestEnrichGPUInfoWithMaxClocks(t *testing.T) {
+func TestEnrichGPUInfoWithNvidiaSMIQ(t *testing.T) {
 	t.Parallel()
 
 	nvsmiQ := []byte(`
 GPU 00000000:4E:00.0
     Product Name                          : NVIDIA RTX PRO 6000 Blackwell Server Edition
+    Min Power Limit                       : 200.00 W
+    Max Power Limit                       : 600.00 W
+    Default Power Limit                   : 575.00 W
+    Current Power Limit                   : 560.00 W
     Clocks
         Graphics                          : 2422 MHz
         Memory                            : 12481 MHz
@@ -365,7 +514,7 @@ GPU 00000000:4F:00.0
 		1: {Index: 1, BusID: "00000000:4F:00.0"},
 	}
 
-	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+	enrichGPUInfoWithNvidiaSMIQ(infoByIndex, nvsmiQ)
 
 	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
 		t.Errorf("GPU 0 MaxGraphicsClockMHz = %v, want 2430", infoByIndex[0].MaxGraphicsClockMHz)
@@ -379,25 +528,49 @@ GPU 00000000:4F:00.0
 	if infoByIndex[1].MaxMemoryClockMHz != 12481 {
 		t.Errorf("GPU 1 MaxMemoryClockMHz = %v, want 12481", infoByIndex[1].MaxMemoryClockMHz)
 	}
+	if infoByIndex[0].MinPowerLimitW != 200 {
+		t.Errorf("GPU 0 MinPowerLimitW = %v, want 200", infoByIndex[0].MinPowerLimitW)
+	}
+	if infoByIndex[0].MaxPowerLimitW != 600 {
+		t.Errorf("GPU 0 MaxPowerLimitW = %v, want 600", infoByIndex[0].MaxPowerLimitW)
+	}
+	if infoByIndex[0].DefaultPowerLimitW != 575 {
+		t.Errorf("GPU 0 DefaultPowerLimitW = %v, want 575", infoByIndex[0].DefaultPowerLimitW)
+	}
+	if infoByIndex[0].PowerLimitW != 560 {
+		t.Errorf("GPU 0 PowerLimitW = %v, want 560", infoByIndex[0].PowerLimitW)
+	}
 }
 
-func TestEnrichGPUInfoWithMaxClocksSkipsPopulated(t *testing.T) {
+func TestEnrichGPUInfoWithNvidiaSMIQSkipsPopulated(t *testing.T) {
 	t.Parallel()
 
 	nvsmiQ := []byte(`
 GPU 00000000:4E:00.0
+    Min Power Limit                       : 100.00 W
+    Max Power Limit                       : 900.00 W
     Max Clocks
         Graphics                          : 9999 MHz
         Memory                            : 9999 MHz
 `)
 	// Already populated — must not be overwritten.
 	infoByIndex := map[int]benchmarkGPUInfo{
-		0: {Index: 0, BusID: "00000000:4E:00.0", MaxGraphicsClockMHz: 2430, MaxMemoryClockMHz: 12481},
+		0: {
+			Index:               0,
+			BusID:               "00000000:4E:00.0",
+			MaxGraphicsClockMHz: 2430,
+			MaxMemoryClockMHz:   12481,
+			MinPowerLimitW:      200,
+			MaxPowerLimitW:      600,
+		},
 	}
 
-	enrichGPUInfoWithMaxClocks(infoByIndex, nvsmiQ)
+	enrichGPUInfoWithNvidiaSMIQ(infoByIndex, nvsmiQ)
 
 	if infoByIndex[0].MaxGraphicsClockMHz != 2430 {
 		t.Errorf("expected existing value to be preserved, got %v", infoByIndex[0].MaxGraphicsClockMHz)
 	}
+	if infoByIndex[0].MinPowerLimitW != 200 {
+		t.Errorf("expected existing min power limit to be preserved, got %v", infoByIndex[0].MinPowerLimitW)
+	}
 }

audit/internal/platform/nvidia_recover.go

@@ -0,0 +1,30 @@
+package platform
+
+import (
+	"fmt"
+	"os/exec"
+	"time"
+)
+
+const nvidiaRecoverHelper = "/usr/local/bin/bee-nvidia-recover"
+
+func runNvidiaRecover(args ...string) (string, error) {
+	helperArgs := append([]string{nvidiaRecoverHelper}, args...)
+	if _, err := exec.LookPath("systemd-run"); err == nil {
+		unit := fmt.Sprintf("bee-nvidia-recover-%d", time.Now().UnixNano())
+		cmdArgs := []string{
+			"systemd-run",
+			"--quiet",
+			"--pipe",
+			"--wait",
+			"--collect",
+			"--service-type=oneshot",
+			"--unit", unit,
+		}
+		cmdArgs = append(cmdArgs, helperArgs...)
+		raw, err := exec.Command("sudo", cmdArgs...).CombinedOutput()
+		return string(raw), err
+	}
+	raw, err := exec.Command("sudo", helperArgs...).CombinedOutput()
+	return string(raw), err
+}

audit/internal/platform/sat.go

@@ -407,11 +407,11 @@ func (s *System) ResetNvidiaGPU(index int) (string, error) {
 	if index < 0 {
 		return "", fmt.Errorf("gpu index must be >= 0")
 	}
-	raw, err := satExecCommand("nvidia-smi", "-r", "-i", strconv.Itoa(index)).CombinedOutput()
-	if len(raw) == 0 && err == nil {
-		raw = []byte("GPU reset completed.\n")
+	out, err := runNvidiaRecover("reset-gpu", strconv.Itoa(index))
+	if strings.TrimSpace(out) == "" && err == nil {
+		out = "GPU reset completed.\n"
 	}
-	return string(raw), err
+	return out, err
 }
 
 // RunNCCLTests runs nccl-tests all_reduce_perf across the selected NVIDIA GPUs.

audit/internal/platform/services.go

@@ -61,6 +61,9 @@ func (s *System) ServiceState(name string) string {
 }
 
 func (s *System) ServiceDo(name string, action ServiceAction) (string, error) {
+	if name == "bee-nvidia" && action == ServiceRestart {
+		return runNvidiaRecover("restart-drivers")
+	}
 	// bee-web runs as the bee user; sudo is required to control system services.
 	// /etc/sudoers.d/bee grants bee NOPASSWD:ALL.
 	raw, err := exec.Command("sudo", "systemctl", string(action), name).CombinedOutput()

iso/overlay/usr/local/bin/bee-nvidia-recover

@@ -0,0 +1,178 @@
+#!/bin/sh
+# bee-nvidia-recover — drain NVIDIA clients, then reset a GPU or reload drivers.
+
+set -u
+
+log() {
+    echo "[bee-nvidia-recover] $*"
+}
+
+log_blocker() {
+    echo "[bee-nvidia-recover] blocker: $*"
+}
+
+usage() {
+    cat <<'EOF'
+usage:
+  bee-nvidia-recover restart-drivers
+  bee-nvidia-recover reset-gpu <index>
+EOF
+}
+
+unit_exists() {
+    systemctl cat "$1" >/dev/null 2>&1
+}
+
+unit_is_active() {
+    systemctl is-active --quiet "$1" 2>/dev/null
+}
+
+stop_unit_if_active() {
+    unit="$1"
+    if unit_is_active "$unit"; then
+        log "stopping $unit"
+        systemctl stop "$unit"
+        return 0
+    fi
+    return 1
+}
+
+start_unit_if_marked() {
+    unit="$1"
+    marker="$2"
+    if [ "$marker" = "1" ] && unit_exists "$unit"; then
+        log "starting $unit"
+        systemctl start "$unit"
+    fi
+}
+
+wait_for_process_exit() {
+    name="$1"
+    tries=0
+    while pgrep -x "$name" >/dev/null 2>&1; do
+        tries=$((tries + 1))
+        if [ "$tries" -ge 15 ]; then
+            log "WARN: $name is still running after stop request"
+            return 1
+        fi
+        sleep 1
+    done
+    return 0
+}
+
+kill_pattern() {
+    pattern="$1"
+    if pgrep -f "$pattern" >/dev/null 2>&1; then
+        pgrep -af "$pattern" 2>/dev/null | while IFS= read -r line; do
+            [ -n "$line" ] || continue
+            log_blocker "$line"
+        done
+        log "killing processes matching: $pattern"
+        pkill -TERM -f "$pattern" >/dev/null 2>&1 || true
+        sleep 1
+        pkill -KILL -f "$pattern" >/dev/null 2>&1 || true
+    fi
+}
+
+drain_gpu_clients() {
+    display_was_active=0
+    fabric_was_active=0
+
+    for unit in display-manager.service lightdm.service; do
+        if unit_exists "$unit" && stop_unit_if_active "$unit"; then
+            log_blocker "service $unit"
+            display_was_active=1
+        fi
+    done
+
+    if unit_exists nvidia-fabricmanager.service && stop_unit_if_active nvidia-fabricmanager.service; then
+        log_blocker "service nvidia-fabricmanager.service"
+        fabric_was_active=1
+    fi
+
+    if pgrep -x nv-hostengine >/dev/null 2>&1; then
+        pgrep -af "^nv-hostengine$" 2>/dev/null | while IFS= read -r line; do
+            [ -n "$line" ] || continue
+            log_blocker "$line"
+        done
+        log "stopping nv-hostengine"
+        pkill -TERM -x nv-hostengine >/dev/null 2>&1 || true
+        wait_for_process_exit nv-hostengine || pkill -KILL -x nv-hostengine >/dev/null 2>&1 || true
+    fi
+
+    for pattern in \
+        "nvidia-smi" \
+        "dcgmi" \
+        "nvvs" \
+        "dcgmproftester" \
+        "all_reduce_perf" \
+        "nvtop" \
+        "bee-gpu-burn" \
+        "bee-john-gpu-stress" \
+        "bee-nccl-gpu-stress" \
+        "Xorg" \
+        "Xwayland"; do
+        kill_pattern "$pattern"
+    done
+}
+
+restore_gpu_clients() {
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        if nvidia-smi -pm 1 >/dev/null 2>&1; then
+            log "enabled NVIDIA persistence mode"
+        else
+            log "WARN: failed to enable NVIDIA persistence mode"
+        fi
+    fi
+
+    if command -v nv-hostengine >/dev/null 2>&1 && ! pgrep -x nv-hostengine >/dev/null 2>&1; then
+        log "starting nv-hostengine"
+        nv-hostengine
+    fi
+
+    start_unit_if_marked nvidia-fabricmanager.service "${fabric_was_active:-0}"
+    start_unit_if_marked display-manager.service "${display_was_active:-0}"
+    if [ "${display_was_active:-0}" = "1" ] && unit_exists lightdm.service && ! unit_is_active lightdm.service; then
+        start_unit_if_marked lightdm.service "1"
+    fi
+}
+
+restart_drivers() {
+    drain_gpu_clients
+    for mod in nvidia_uvm nvidia_drm nvidia_modeset nvidia; do
+        if lsmod | awk '{print $1}' | grep -qx "$mod"; then
+            log "unloading module $mod"
+            rmmod "$mod"
+        fi
+    done
+    rm -f /dev/nvidiactl /dev/nvidia-uvm /dev/nvidia-uvm-tools /dev/nvidia[0-9]* 2>/dev/null || true
+    log "reloading NVIDIA driver stack"
+    /usr/local/bin/bee-nvidia-load
+    restore_gpu_clients
+}
+
+reset_gpu() {
+    index="$1"
+    drain_gpu_clients
+    log "resetting GPU $index"
+    nvidia-smi -r -i "$index"
+    restore_gpu_clients
+}
+
+cmd="${1:-}"
+case "$cmd" in
+    restart-drivers)
+        restart_drivers
+        ;;
+    reset-gpu)
+        if [ "$#" -ne 2 ]; then
+            usage >&2
+            exit 2
+        fi
+        reset_gpu "$2"
+        ;;
+    *)
+        usage >&2
+        exit 2
+        ;;
+esac