Seed power ramp with single-card TDP limits

audit/internal/app/app.go

@@ -146,7 +146,7 @@ type satRunner interface {
 	RunSATStressPack(ctx context.Context, baseDir string, durationSec int, logFunc func(string)) (string, error)
 	RunFanStressTest(ctx context.Context, baseDir string, opts platform.FanStressOptions) (string, error)
 	RunPlatformStress(ctx context.Context, baseDir string, opts platform.PlatformStressOptions, logFunc func(string)) (string, error)
-	RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error)
+	RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error)
 }
 
 type runtimeChecker interface {
@@ -744,8 +744,15 @@ func (a *App) RunPlatformStress(ctx context.Context, baseDir string, opts platfo
 	return a.sat.RunPlatformStress(ctx, baseDir, opts, logFunc)
 }
 
+func (a *App) RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	if strings.TrimSpace(baseDir) == "" {
+		baseDir = DefaultSATBaseDir
+	}
+	return a.sat.RunNCCLTests(ctx, baseDir, gpuIndices, logFunc)
+}
+
 func (a *App) RunNCCLTestsResult(ctx context.Context) (ActionResult, error) {
-	path, err := a.sat.RunNCCLTests(ctx, DefaultSATBaseDir, nil)
+	path, err := a.RunNCCLTests(ctx, DefaultSATBaseDir, nil, nil)
 	body := "Results: " + path
 	if err != nil && err != context.Canceled {
 		body += "\nERROR: " + err.Error()

audit/internal/app/app_test.go

@@ -128,6 +128,7 @@ type fakeSAT struct {
 	runNvidiaPowerFn          func(string, int, []int) (string, error)
 	runNvidiaPulseFn          func(string, int, []int) (string, error)
 	runNvidiaBandwidthFn      func(string, []int) (string, error)
+	runNCCLFn                 func(string, []int) (string, error)
 	runNvidiaTargetedStressFn func(string, int, []int) (string, error)
 	runMemoryFn               func(string) (string, error)
 	runStorageFn              func(string) (string, error)
@@ -287,10 +288,43 @@ func (f fakeSAT) RunPlatformStress(_ context.Context, _ string, _ platform.Platf
 	return "", nil
 }
 
-func (f fakeSAT) RunNCCLTests(_ context.Context, _ string, _ func(string)) (string, error) {
+func (f fakeSAT) RunNCCLTests(_ context.Context, baseDir string, gpuIndices []int, _ func(string)) (string, error) {
+	if f.runNCCLFn != nil {
+		return f.runNCCLFn(baseDir, gpuIndices)
+	}
 	return "", nil
 }
 
+func TestRunNCCLTestsPassesSelectedGPUs(t *testing.T) {
+	t.Parallel()
+
+	var gotBaseDir string
+	var gotGPUIndices []int
+	a := &App{
+		sat: fakeSAT{
+			runNCCLFn: func(baseDir string, gpuIndices []int) (string, error) {
+				gotBaseDir = baseDir
+				gotGPUIndices = append([]int(nil), gpuIndices...)
+				return "/tmp/nccl-tests.tar.gz", nil
+			},
+		},
+	}
+
+	path, err := a.RunNCCLTests(context.Background(), "/tmp/sat", []int{3, 1}, nil)
+	if err != nil {
+		t.Fatalf("RunNCCLTests error: %v", err)
+	}
+	if path != "/tmp/nccl-tests.tar.gz" {
+		t.Fatalf("path=%q want %q", path, "/tmp/nccl-tests.tar.gz")
+	}
+	if gotBaseDir != "/tmp/sat" {
+		t.Fatalf("baseDir=%q want %q", gotBaseDir, "/tmp/sat")
+	}
+	if len(gotGPUIndices) != 2 || gotGPUIndices[0] != 3 || gotGPUIndices[1] != 1 {
+		t.Fatalf("gpuIndices=%v want [3 1]", gotGPUIndices)
+	}
+}
+
 func TestNetworkStatusFormatsInterfacesAndRoute(t *testing.T) {
 	t.Parallel()
 

audit/internal/platform/benchmark.go

@@ -1122,6 +1122,7 @@ type benchmarkCoolingSample struct {
 	AvgFanRPM             float64
 	AvgFanDutyCyclePct    float64
 	FanDutyCycleAvailable bool
+	FanDutyCycleEstimated bool
 }
 
 func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
@@ -1134,6 +1135,7 @@ func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
 		samples[i].FanAvgRPM = fanSample.AvgFanRPM
 		samples[i].FanDutyCyclePct = fanSample.AvgFanDutyCyclePct
 		samples[i].FanDutyCycleAvailable = fanSample.FanDutyCycleAvailable
+		samples[i].FanDutyCycleEstimated = fanSample.FanDutyCycleEstimated
 	}
 	return samples, nil
 }
@@ -1141,11 +1143,12 @@ func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
 func sampleBenchmarkCoolingSample() benchmarkCoolingSample {
 	fans, _ := sampleFanSpeeds()
 	avgRPM, _, _ := fanRPMStats(fans)
-	dutyPct, dutyAvailable := sampleFanDutyCyclePct()
+	dutyPct, dutyAvailable, dutyEstimated := sampleFanDutyCyclePctFromFans(fans)
 	return benchmarkCoolingSample{
 		AvgFanRPM:             avgRPM,
 		AvgFanDutyCyclePct:    dutyPct,
 		FanDutyCycleAvailable: dutyAvailable,
+		FanDutyCycleEstimated: dutyEstimated,
 	}
 }
 
@@ -1387,25 +1390,33 @@ func summarizeBenchmarkCooling(rows []GPUMetricRow) *BenchmarkCoolingSummary {
 	}
 	var rpmValues []float64
 	var dutyValues []float64
+	var dutyEstimated bool
 	for _, row := range rows {
 		if row.FanAvgRPM > 0 {
 			rpmValues = append(rpmValues, row.FanAvgRPM)
 		}
 		if row.FanDutyCycleAvailable {
 			dutyValues = append(dutyValues, row.FanDutyCyclePct)
+			if row.FanDutyCycleEstimated {
+				dutyEstimated = true
+			}
 		}
 	}
 	if len(rpmValues) == 0 && len(dutyValues) == 0 {
 		return nil
 	}
 	summary := &BenchmarkCoolingSummary{
-		Available: true,
-		AvgFanRPM: benchmarkMean(rpmValues),
+		Available:             true,
+		AvgFanRPM:             benchmarkMean(rpmValues),
+		FanDutyCycleEstimated: dutyEstimated,
 	}
 	if len(dutyValues) > 0 {
 		summary.FanDutyCycleAvailable = true
 		summary.AvgFanDutyCyclePct = benchmarkMean(dutyValues)
 		summary.P95FanDutyCyclePct = benchmarkPercentile(dutyValues, 95)
+		if summary.FanDutyCycleEstimated {
+			summary.Notes = append(summary.Notes, "fan duty cycle is estimated from the highest fan RPM observed since boot; treat it as an approximation, not a direct PWM reading")
+		}
 	} else {
 		summary.Notes = append(summary.Notes, "fan duty cycle unavailable on this host; RPM-only fan telemetry was collected")
 	}
@@ -3491,10 +3502,21 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		_ = os.MkdirAll(stepDir, 0755)
 
 		// Reuse the latest stable limits as starting points, but re-check every
-		// active GPU in this hotter configuration.
-		seedForStep := make(map[int]int, len(stableLimits))
-		for k, v := range stableLimits {
-			seedForStep[k] = v
+		// active GPU in this hotter configuration. For the newly introduced GPU,
+		// seed from its single-card calibration so we do not restart from the
+		// default TDP when a prior derated limit is already known.
+		seedForStep := make(map[int]int, len(subset))
+		for _, idx := range subset {
+			if lim, ok := stableLimits[idx]; ok && lim > 0 {
+				seedForStep[idx] = lim
+				continue
+			}
+			if base, ok := calibByIndex[idx]; ok {
+				lim := int(math.Round(base.AppliedPowerLimitW))
+				if lim > 0 {
+					seedForStep[idx] = lim
+				}
+			}
 		}
 
 		logFunc(fmt.Sprintf("power ramp: step %d/%d — revalidating %d active GPU(s) including new GPU %d",

audit/internal/platform/benchmark_types.go

@@ -31,6 +31,7 @@ type BenchmarkCoolingSummary struct {
 	Available             bool     `json:"available"`
 	AvgFanRPM             float64  `json:"avg_fan_rpm,omitempty"`
 	FanDutyCycleAvailable bool     `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool     `json:"fan_duty_cycle_estimated,omitempty"`
 	AvgFanDutyCyclePct    float64  `json:"avg_fan_duty_cycle_pct,omitempty"`
 	P95FanDutyCyclePct    float64  `json:"p95_fan_duty_cycle_pct,omitempty"`
 	Notes                 []string `json:"notes,omitempty"`
@@ -55,32 +56,32 @@ type NvidiaBenchmarkOptions struct {
 }
 
 type NvidiaBenchmarkResult struct {
-	BenchmarkVersion   string                       `json:"benchmark_version"`
-	GeneratedAt        time.Time                    `json:"generated_at"`
-	Hostname           string                       `json:"hostname,omitempty"`
-	ServerModel        string                       `json:"server_model,omitempty"`
-	BenchmarkProfile   string                       `json:"benchmark_profile"`
-	ParallelGPUs       bool                         `json:"parallel_gpus,omitempty"`
-	RampStep           int                          `json:"ramp_step,omitempty"`
-	RampTotal          int                          `json:"ramp_total,omitempty"`
-	RampRunID          string                       `json:"ramp_run_id,omitempty"`
-	ScalabilityScore   float64                      `json:"scalability_score,omitempty"`
+	BenchmarkVersion string    `json:"benchmark_version"`
+	GeneratedAt      time.Time `json:"generated_at"`
+	Hostname         string    `json:"hostname,omitempty"`
+	ServerModel      string    `json:"server_model,omitempty"`
+	BenchmarkProfile string    `json:"benchmark_profile"`
+	ParallelGPUs     bool      `json:"parallel_gpus,omitempty"`
+	RampStep         int       `json:"ramp_step,omitempty"`
+	RampTotal        int       `json:"ramp_total,omitempty"`
+	RampRunID        string    `json:"ramp_run_id,omitempty"`
+	ScalabilityScore float64   `json:"scalability_score,omitempty"`
 	// PlatformPowerScore is the mean compute scalability across ramp steps 2..N.
 	// 100% = each added GPU contributes exactly its single-card throughput.
 	// < 100% = throughput loss due to thermal throttle, power limits, or contention.
-	PlatformPowerScore   float64                    `json:"platform_power_score,omitempty"`
-	PerformanceRampSteps []NvidiaPerformanceRampStep `json:"performance_ramp_steps,omitempty"`
-	OverallStatus      string                       `json:"overall_status"`
-	SelectedGPUIndices []int                        `json:"selected_gpu_indices"`
-	Findings           []string                     `json:"findings,omitempty"`
-	Warnings           []string                     `json:"warnings,omitempty"`
-	Normalization      BenchmarkNormalization       `json:"normalization"`
-	HostConfig         *BenchmarkHostConfig         `json:"host_config,omitempty"`
-	CPULoad            *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
-	Cooling            *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
-	GPUs               []BenchmarkGPUResult         `json:"gpus"`
-	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
-	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`
+	PlatformPowerScore   float64                      `json:"platform_power_score,omitempty"`
+	PerformanceRampSteps []NvidiaPerformanceRampStep  `json:"performance_ramp_steps,omitempty"`
+	OverallStatus        string                       `json:"overall_status"`
+	SelectedGPUIndices   []int                        `json:"selected_gpu_indices"`
+	Findings             []string                     `json:"findings,omitempty"`
+	Warnings             []string                     `json:"warnings,omitempty"`
+	Normalization        BenchmarkNormalization       `json:"normalization"`
+	HostConfig           *BenchmarkHostConfig         `json:"host_config,omitempty"`
+	CPULoad              *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
+	Cooling              *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
+	GPUs                 []BenchmarkGPUResult         `json:"gpus"`
+	Interconnect         *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
+	ServerPower          *BenchmarkServerPower        `json:"server_power,omitempty"`
 }
 
 type BenchmarkNormalization struct {
@@ -223,8 +224,8 @@ type BenchmarkScorecard struct {
 
 	// Throttle breakdown — percentage of steady-state time in each throttle type.
 	// Used for diagnosis: tells WHY the GPU throttled, not just whether it did.
-	ThermalThrottlePct  float64 `json:"thermal_throttle_pct"`  // HW+SW thermal slowdown
-	PowerCapThrottlePct float64 `json:"power_cap_throttle_pct"` // SW power cap
+	ThermalThrottlePct   float64 `json:"thermal_throttle_pct"`   // HW+SW thermal slowdown
+	PowerCapThrottlePct  float64 `json:"power_cap_throttle_pct"` // SW power cap
 	SyncBoostThrottlePct float64 `json:"sync_boost_throttle_pct,omitempty"`
 
 	// Temperature headroom: distance to the 100°C destruction threshold.
@@ -300,22 +301,22 @@ type NvidiaPowerBenchResult struct {
 	// PlatformMaxTDPW is the sum of per-GPU stable power limits found during the
 	// cumulative thermal ramp. Represents the actual sustained power budget of
 	// this server under full GPU load. Use for rack power planning.
-	PlatformMaxTDPW float64               `json:"platform_max_tdp_w"`
+	PlatformMaxTDPW float64 `json:"platform_max_tdp_w"`
 	// ServerPower captures IPMI server power delta (idle→loaded) measured in
 	// parallel with the thermal ramp. Use to compare GPU-reported TDP against
 	// actual wall-power draw as seen by the server's power supply.
-	ServerPower     *BenchmarkServerPower `json:"server_power,omitempty"`
-	Findings        []string              `json:"findings,omitempty"`
-	GPUs            []NvidiaPowerBenchGPU `json:"gpus"`
+	ServerPower *BenchmarkServerPower `json:"server_power,omitempty"`
+	Findings    []string              `json:"findings,omitempty"`
+	GPUs        []NvidiaPowerBenchGPU `json:"gpus"`
 }
 
 type NvidiaPowerBenchGPU struct {
-	Index               int      `json:"index"`
-	Name                string   `json:"name,omitempty"`
-	BusID               string   `json:"bus_id,omitempty"`
-	DefaultPowerLimitW  float64  `json:"default_power_limit_w,omitempty"`
+	Index              int     `json:"index"`
+	Name               string  `json:"name,omitempty"`
+	BusID              string  `json:"bus_id,omitempty"`
+	DefaultPowerLimitW float64 `json:"default_power_limit_w,omitempty"`
 	// AppliedPowerLimitW is the stable limit found during single-card calibration.
-	AppliedPowerLimitW  float64  `json:"applied_power_limit_w,omitempty"`
+	AppliedPowerLimitW float64 `json:"applied_power_limit_w,omitempty"`
 	// StablePowerLimitW is the final fixed limit for this GPU after the
 	// cumulative thermal ramp. This is the limit at which the GPU operated
 	// stably with all other GPUs running simultaneously at their own limits.
@@ -333,10 +334,10 @@ type NvidiaPowerBenchGPU struct {
 }
 
 type NvidiaPowerBenchStep struct {
-	StepIndex           int      `json:"step_index"`
-	GPUIndices          []int    `json:"gpu_indices"`
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
 	// NewGPUIndex is the GPU whose stable limit was searched in this step.
-	NewGPUIndex         int      `json:"new_gpu_index"`
+	NewGPUIndex int `json:"new_gpu_index"`
 	// NewGPUStableLimitW is the stable power limit found for the new GPU.
 	NewGPUStableLimitW  float64  `json:"new_gpu_stable_limit_w,omitempty"`
 	TotalObservedPowerW float64  `json:"total_observed_power_w,omitempty"`
@@ -349,15 +350,15 @@ type NvidiaPowerBenchStep struct {
 // NvidiaPerformanceRampStep holds per-step performance data for the
 // scalability ramp-up phase of the performance benchmark.
 type NvidiaPerformanceRampStep struct {
-	StepIndex          int      `json:"step_index"`
-	GPUIndices         []int    `json:"gpu_indices"`
+	StepIndex  int   `json:"step_index"`
+	GPUIndices []int `json:"gpu_indices"`
 	// TotalSyntheticTOPS is the sum of per-GPU SyntheticScore (fp32-equivalent
 	// TOPS from dedicated single-precision phases) across all GPUs in this step.
-	TotalSyntheticTOPS float64  `json:"total_synthetic_tops"`
-	TotalMixedTOPS     float64  `json:"total_mixed_tops,omitempty"`
+	TotalSyntheticTOPS float64 `json:"total_synthetic_tops"`
+	TotalMixedTOPS     float64 `json:"total_mixed_tops,omitempty"`
 	// ScalabilityPct = TotalSyntheticTOPS / (k × best_single_gpu_tops) × 100.
 	// 100% = perfect linear scaling. < 100% = thermal/power/interconnect loss.
-	ScalabilityPct     float64  `json:"scalability_pct"`
-	Status             string   `json:"status"`
-	Notes              []string `json:"notes,omitempty"`
+	ScalabilityPct float64  `json:"scalability_pct"`
+	Status         string   `json:"status"`
+	Notes          []string `json:"notes,omitempty"`
 }

audit/internal/platform/gpu_metrics.go

@@ -27,6 +27,7 @@ type GPUMetricRow struct {
 	FanAvgRPM             float64 `json:"fan_avg_rpm,omitempty"`
 	FanDutyCyclePct       float64 `json:"fan_duty_cycle_pct,omitempty"`
 	FanDutyCycleAvailable bool    `json:"fan_duty_cycle_available,omitempty"`
+	FanDutyCycleEstimated bool    `json:"fan_duty_cycle_estimated,omitempty"`
 }
 
 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
@@ -147,14 +148,18 @@ func sampleAMDGPUMetrics() ([]GPUMetricRow, error) {
 // WriteGPUMetricsCSV writes collected rows as a CSV file.
 func WriteGPUMetricsCSV(path string, rows []GPUMetricRow) error {
 	var b bytes.Buffer
-	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz,fan_avg_rpm,fan_duty_cycle_pct,fan_duty_cycle_available\n")
+	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz,fan_avg_rpm,fan_duty_cycle_pct,fan_duty_cycle_available,fan_duty_cycle_estimated\n")
 	for _, r := range rows {
 		dutyAvail := 0
 		if r.FanDutyCycleAvailable {
 			dutyAvail = 1
 		}
-		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%d\n",
-			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz, r.FanAvgRPM, r.FanDutyCyclePct, dutyAvail)
+		dutyEstimated := 0
+		if r.FanDutyCycleEstimated {
+			dutyEstimated = 1
+		}
+		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%d,%d\n",
+			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz, r.FanAvgRPM, r.FanDutyCyclePct, dutyAvail, dutyEstimated)
 	}
 	return os.WriteFile(path, b.Bytes(), 0644)
 }

audit/internal/platform/sat.go

@@ -366,12 +366,14 @@ func (s *System) ResetNvidiaGPU(index int) (string, error) {
 	return string(raw), err
 }
 
-// RunNCCLTests runs nccl-tests all_reduce_perf across all NVIDIA GPUs.
+// RunNCCLTests runs nccl-tests all_reduce_perf across the selected NVIDIA GPUs.
 // Measures collective communication bandwidth over NVLink/PCIe.
-func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(string)) (string, error) {
-	// detect GPU count
-	out, _ := exec.Command("nvidia-smi", "--query-gpu=index", "--format=csv,noheader").Output()
-	gpuCount := len(strings.Split(strings.TrimSpace(string(out)), "\n"))
+func (s *System) RunNCCLTests(ctx context.Context, baseDir string, gpuIndices []int, logFunc func(string)) (string, error) {
+	selected, err := resolveDCGMGPUIndices(gpuIndices)
+	if err != nil {
+		return "", err
+	}
+	gpuCount := len(selected)
 	if gpuCount < 1 {
 		gpuCount = 1
 	}
@@ -380,7 +382,7 @@ func (s *System) RunNCCLTests(ctx context.Context, baseDir string, logFunc func(
 		satJob{name: "02-all-reduce-perf.log", cmd: []string{
 			"all_reduce_perf", "-b", "512M", "-e", "4G", "-f", "2",
 			"-g", strconv.Itoa(gpuCount), "--iters", "20",
-		}},
+		}, env: nvidiaVisibleDevicesEnv(selected)},
 	), logFunc)
 }
 

audit/internal/platform/sat_fan_stress.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"math"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -56,13 +57,37 @@ type cachedPowerReading struct {
 	UpdatedAt time.Time
 }
 
+type fanObservationState struct {
+	MaxRPM map[string]float64 `json:"max_rpm"`
+}
+
+type fanPeakCandidate struct {
+	FirstSeen time.Time
+	RPM       float64
+}
+
 var (
 	systemPowerCacheMu sync.Mutex
 	systemPowerCache   cachedPowerReading
+	fanObservationMu   sync.Mutex
+	fanObservation     fanObservationState
+	fanObservationInit bool
+	fanPeakCandidates  = make(map[string]fanPeakCandidate)
 )
 
 const systemPowerHoldTTL = 15 * time.Second
 
+var fanObservationStatePath = "/var/log/bee-sat/fan-observation.json"
+
+const fanObservationMinPeakHold = time.Second
+
+func normalizeObservedFanMaxRPM(rpm float64) float64 {
+	if rpm <= 0 {
+		return 0
+	}
+	return math.Ceil(rpm/1000.0) * 1000.0
+}
+
 // RunFanStressTest runs a two-phase GPU stress test while monitoring fan speeds,
 // temperatures, and power draw every second. Exports metrics.csv and fan-sensors.csv.
 // Designed to reproduce case-04 fan-speed lag and detect GPU thermal throttling.
@@ -310,11 +335,13 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	out, err := exec.Command("ipmitool", "sdr", "type", "Fan").Output()
 	if err == nil {
 		if fans := parseFanSpeeds(string(out)); len(fans) > 0 {
+			updateFanObservation(fans, time.Now())
 			return fans, nil
 		}
 	}
 	fans, sensorsErr := sampleFanSpeedsViaSensorsJSON()
 	if len(fans) > 0 {
+		updateFanObservation(fans, time.Now())
 		return fans, nil
 	}
 	if err != nil {
@@ -323,6 +350,119 @@ func sampleFanSpeeds() ([]FanReading, error) {
 	return nil, sensorsErr
 }
 
+func loadFanObservationLocked() {
+	if fanObservationInit {
+		return
+	}
+	fanObservationInit = true
+	fanObservation.MaxRPM = make(map[string]float64)
+	raw, err := os.ReadFile(fanObservationStatePath)
+	if err != nil || len(raw) == 0 {
+		return
+	}
+	var persisted fanObservationState
+	if json.Unmarshal(raw, &persisted) != nil {
+		return
+	}
+	for name, rpm := range persisted.MaxRPM {
+		name = strings.TrimSpace(name)
+		if name == "" || rpm <= 0 {
+			continue
+		}
+		fanObservation.MaxRPM[name] = rpm
+	}
+}
+
+func saveFanObservationLocked() {
+	if len(fanObservation.MaxRPM) == 0 {
+		return
+	}
+	dir := filepath.Dir(fanObservationStatePath)
+	if dir == "" || dir == "." {
+		dir = "/var/log/bee-sat"
+	}
+	if err := os.MkdirAll(dir, 0755); err != nil {
+		return
+	}
+	raw, err := json.MarshalIndent(fanObservation, "", "  ")
+	if err != nil {
+		return
+	}
+	_ = os.WriteFile(fanObservationStatePath, raw, 0644)
+}
+
+func updateFanObservation(fans []FanReading, now time.Time) {
+	if len(fans) == 0 {
+		return
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	changed := false
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		currentMax := fanObservation.MaxRPM[name]
+		if fan.RPM <= currentMax {
+			delete(fanPeakCandidates, name)
+			continue
+		}
+		if cand, ok := fanPeakCandidates[name]; ok {
+			if now.Sub(cand.FirstSeen) >= fanObservationMinPeakHold {
+				newMax := math.Max(cand.RPM, fan.RPM)
+				if newMax > currentMax {
+					fanObservation.MaxRPM[name] = normalizeObservedFanMaxRPM(newMax)
+					changed = true
+				}
+				delete(fanPeakCandidates, name)
+				continue
+			}
+			if fan.RPM > cand.RPM {
+				fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: cand.FirstSeen, RPM: fan.RPM}
+			}
+			continue
+		}
+		fanPeakCandidates[name] = fanPeakCandidate{FirstSeen: now, RPM: fan.RPM}
+	}
+	if changed {
+		saveFanObservationLocked()
+	}
+}
+
+func estimateFanDutyCyclePctFromObservation(fans []FanReading) (float64, bool) {
+	if len(fans) == 0 {
+		return 0, false
+	}
+	fanObservationMu.Lock()
+	defer fanObservationMu.Unlock()
+	loadFanObservationLocked()
+	var samples []float64
+	for _, fan := range fans {
+		name := strings.TrimSpace(fan.Name)
+		if name == "" || fan.RPM <= 0 {
+			continue
+		}
+		maxRPM := fanObservation.MaxRPM[name]
+		if maxRPM <= 0 {
+			continue
+		}
+		pct := fan.RPM / maxRPM * 100.0
+		if pct > 100 {
+			pct = 100
+		}
+		if pct < 0 {
+			pct = 0
+		}
+		samples = append(samples, pct)
+	}
+	if len(samples) == 0 {
+		return 0, false
+	}
+	return benchmarkMean(samples), true
+}
+
 // parseFanSpeeds parses "ipmitool sdr type Fan" output.
 // Handles two formats:
 //
@@ -428,12 +568,27 @@ func sampleFanSpeedsViaSensorsJSON() ([]FanReading, error) {
 
 // sampleFanDutyCyclePct reads fan PWM/duty-cycle controls from lm-sensors.
 // Returns the average duty cycle across all exposed PWM controls.
-func sampleFanDutyCyclePct() (float64, bool) {
+func sampleFanDutyCyclePct() (float64, bool, bool) {
 	out, err := exec.Command("sensors", "-j").Output()
 	if err != nil || len(out) == 0 {
-		return 0, false
+		fans, fanErr := sampleFanSpeeds()
+		if fanErr != nil {
+			return 0, false, false
+		}
+		return sampleFanDutyCyclePctFromFans(fans)
 	}
-	return parseFanDutyCyclePctSensorsJSON(out)
+	pct, ok := parseFanDutyCyclePctSensorsJSON(out)
+	return pct, ok, false
+}
+
+func sampleFanDutyCyclePctFromFans(fans []FanReading) (float64, bool, bool) {
+	if len(fans) == 0 {
+		return 0, false, false
+	}
+	if pct, ok := estimateFanDutyCyclePctFromObservation(fans); ok {
+		return pct, true, true
+	}
+	return 0, false, false
 }
 
 func parseFanDutyCyclePctSensorsJSON(raw []byte) (float64, bool) {

audit/internal/platform/sat_fan_stress_test.go

@@ -1,6 +1,7 @@
 package platform
 
 import (
+	"path/filepath"
 	"testing"
 	"time"
 )
@@ -50,6 +51,53 @@ func TestParseFanDutyCyclePctSensorsJSON(t *testing.T) {
 	}
 }
 
+func TestEstimateFanDutyCyclePctFromObservation(t *testing.T) {
+	t.Parallel()
+
+	oldPath := fanObservationStatePath
+	oldState := fanObservation
+	oldInit := fanObservationInit
+	oldCandidates := fanPeakCandidates
+	fanObservationStatePath = filepath.Join(t.TempDir(), "fan-observation.json")
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	t.Cleanup(func() {
+		fanObservationStatePath = oldPath
+		fanObservation = oldState
+		fanObservationInit = oldInit
+		fanPeakCandidates = oldCandidates
+	})
+
+	start := time.Unix(100, 0)
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5000}}, start)
+	if _, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2500}}); ok {
+		t.Fatalf("single-sample spike should not establish observed max")
+	}
+
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5200}}, start.Add(500*time.Millisecond))
+	updateFanObservation([]FanReading{{Name: "FAN1", RPM: 5100}}, start.Add(1500*time.Millisecond))
+
+	got, ok := estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected estimated duty cycle from persisted observed max")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("got=%v want ~43.3", got)
+	}
+
+	fanObservation = fanObservationState{}
+	fanObservationInit = false
+	fanPeakCandidates = make(map[string]fanPeakCandidate)
+	got, ok = estimateFanDutyCyclePctFromObservation([]FanReading{{Name: "FAN1", RPM: 2600}})
+	if !ok {
+		t.Fatalf("expected persisted observed max to be reloaded from disk")
+	}
+	if got < 43 || got > 44 {
+		t.Fatalf("reloaded got=%v want ~43.3", got)
+	}
+}
+
 func TestParseDCMIPowerReading(t *testing.T) {
 	raw := `
 Instantaneous power reading:                   512 Watts

audit/internal/platform/sat_test.go

@@ -321,6 +321,19 @@ func TestNvidiaDCGMNamedDiagCommandUsesDurationAndSelection(t *testing.T) {
 	}
 }
 
+func TestNvidiaDCGMNamedDiagCommandSkipsDurationForNVBandwidth(t *testing.T) {
+	cmd := nvidiaDCGMNamedDiagCommand("nvbandwidth", 0, []int{2, 0})
+	want := []string{"dcgmi", "diag", "-r", "nvbandwidth", "-i", "2,0"}
+	if len(cmd) != len(want) {
+		t.Fatalf("cmd len=%d want %d (%v)", len(cmd), len(want), cmd)
+	}
+	for i := range want {
+		if cmd[i] != want[i] {
+			t.Fatalf("cmd[%d]=%q want %q", i, cmd[i], want[i])
+		}
+	}
+}
+
 func TestNvidiaVisibleDevicesEnvUsesSelectedGPUs(t *testing.T) {
 	env := nvidiaVisibleDevicesEnv([]int{0, 2, 4})
 	if len(env) != 2 {

audit/internal/webui/pages.go

@@ -1481,7 +1481,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Verifies NVLink/NVSwitch fabric bandwidth using NCCL all_reduce_perf across all selected GPUs. Pass/fail based on achieved bandwidth vs. theoretical.`,
 			`<code>all_reduce_perf</code> (NCCL tests)`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs across all selected GPUs simultaneously (requires ≥2).<p id="sat-ni-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			`Runs in Validate and Stress. Uses all selected GPUs simultaneously (requires ≥2) and is kept short so it fits the Validate flow.`,
 		)) +
 		`</div>` +
 		`<div id="sat-card-nvidia-bandwidth">` +
@@ -1489,7 +1489,7 @@ func renderValidate(opts HandlerOptions) string {
 			inv.NVIDIA,
 			`Validates GPU memory copy and peer-to-peer bandwidth paths using NVBandwidth.`,
 			`<code>nvbandwidth</code>`,
-			`Skipped in Validate mode. Runs in Stress mode only. Runs across all selected GPUs simultaneously.<p id="sat-nb-mode-hint" style="color:var(--warn-fg);font-size:12px;margin:8px 0 0">Only runs in Stress mode. Switch mode above to enable in Run All.</p>`,
+			`Runs in Validate and Stress across all selected GPUs simultaneously. Intended to stay short enough for Validate.`,
 		)) +
 		`</div>` +
 		`</div>
@@ -1527,8 +1527,6 @@ function satModeChanged() {
     {card: 'sat-card-nvidia-targeted-stress', hint: 'sat-ts-mode-hint'},
     {card: 'sat-card-nvidia-targeted-power',  hint: 'sat-tp-mode-hint'},
     {card: 'sat-card-nvidia-pulse',           hint: 'sat-pt-mode-hint'},
-    {card: 'sat-card-nvidia-interconnect',    hint: 'sat-ni-mode-hint'},
-    {card: 'sat-card-nvidia-bandwidth',       hint: 'sat-nb-mode-hint'},
   ].forEach(function(item) {
     const card = document.getElementById(item.card);
     if (card) {
@@ -1776,7 +1774,7 @@ function runAllSAT() {
   const cycles = 1;
   const status = document.getElementById('sat-all-status');
   status.textContent = 'Enqueuing...';
-  const stressOnlyTargets = ['nvidia-targeted-stress', 'nvidia-targeted-power', 'nvidia-pulse', 'nvidia-interconnect', 'nvidia-bandwidth'];
+  const stressOnlyTargets = ['nvidia-targeted-stress', 'nvidia-targeted-power', 'nvidia-pulse'];
   const baseTargets = ['nvidia','nvidia-targeted-stress','nvidia-targeted-power','nvidia-pulse','nvidia-interconnect','nvidia-bandwidth','memory','storage','cpu'].concat(selectedAMDValidateTargets());
   const activeTargets = baseTargets.filter(target => {
     if (stressOnlyTargets.indexOf(target) >= 0 && !satStressMode()) return false;
@@ -2082,7 +2080,7 @@ func renderBenchmark(opts HandlerOptions) string {
   </div>
 </div>
 
-`+`<div id="benchmark-results-section">`+renderBenchmarkResultsCard(opts.ExportDir)+`</div>`+`
+` + `<div id="benchmark-results-section">` + renderBenchmarkResultsCard(opts.ExportDir) + `</div>` + `
 
 <div id="benchmark-output" style="display:none;margin-top:16px" class="card">
   <div class="card-head">Benchmark Output <span id="benchmark-title"></span></div>
@@ -2517,7 +2515,7 @@ func renderPowerBenchmarkResultsCard(exportDir string) string {
 
 func renderBurn() string {
 	return `<div class="alert alert-warn" style="margin-bottom:16px"><strong>&#9888; Warning:</strong> Stress tests on this page run hardware at high load. Repeated or prolonged use may reduce hardware lifespan. Use only when necessary.</div>
-<div class="alert alert-info" style="margin-bottom:16px"><strong>Scope:</strong> DCGM diagnostics (` + "targeted_stress, targeted_power, pulse_test" + `), NCCL, NVBandwidth, and LINPACK remain in <a href="/validate">Validate → Stress mode</a>. Burn exposes sustained GPU compute load recipes.</div>
+<div class="alert alert-info" style="margin-bottom:16px"><strong>Scope:</strong> Burn exposes sustained GPU compute load recipes. DCGM diagnostics (` + "targeted_stress, targeted_power, pulse_test" + `) and LINPACK remain in <a href="/validate">Validate → Stress mode</a>; NCCL and NVBandwidth are available directly from <a href="/validate">Validate</a>.</div>
 <p style="color:var(--muted);font-size:13px;margin-bottom:16px">Tasks continue in the background — view progress in <a href="/tasks">Tasks</a>.</p>
 
 <div class="card" style="margin-bottom:16px">

audit/internal/webui/server_test.go

@@ -744,6 +744,26 @@ func TestValidatePageRendersNvidiaTargetedStressCard(t *testing.T) {
 	}
 }
 
+func TestValidatePageRendersNvidiaFabricCardsInValidateMode(t *testing.T) {
+	handler := NewHandler(HandlerOptions{})
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, "/validate", nil))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d", rec.Code)
+	}
+	body := rec.Body.String()
+	for _, needle := range []string{
+		`NVIDIA Interconnect (NCCL)`,
+		`Runs in Validate and Stress.`,
+		`NVIDIA Bandwidth (NVBandwidth)`,
+		`Intended to stay short enough for Validate.`,
+	} {
+		if !strings.Contains(body, needle) {
+			t.Fatalf("validate page missing %q: %s", needle, body)
+		}
+	}
+}
+
 func TestBurnPageRendersGoalBasedNVIDIACards(t *testing.T) {
 	handler := NewHandler(HandlerOptions{})
 	rec := httptest.NewRecorder()

audit/internal/webui/tasks.go

@@ -736,15 +736,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			err = fmt.Errorf("app not configured")
 			break
 		}
-		dur := t.params.Duration
-		if t.params.BurnProfile != "" && dur <= 0 {
-			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
-		}
-		archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
-			DurationSec: dur,
-			Loader:      platform.NvidiaStressLoaderNCCL,
-			GPUIndices:  t.params.GPUIndices,
-		}, j.append)
+		archive, err = a.RunNCCLTests(ctx, "", t.params.GPUIndices, j.append)
 	case "nvidia-stress":
 		if a == nil {
 			err = fmt.Errorf("app not configured")

iso/builder/config/hooks/normal/9001-wallpaper.hook.chroot

@@ -1,117 +0,0 @@
-#!/bin/sh
-# 9001-wallpaper.hook.chroot — generate /usr/share/bee/wallpaper.png inside chroot
-set -e
-echo "=== generating bee wallpaper ==="
-mkdir -p /usr/share/bee
-
-python3 - <<'PYEOF'
-from PIL import Image, ImageDraw, ImageFont, ImageFilter
-import os
-
-W, H = 1920, 1080
-
-ASCII_ART = [
-    "  ███████╗ █████╗ ███████╗██╗   ██╗      ██████╗ ███████╗███████╗",
-    "  ██╔════╝██╔══██╗██╔════╝╚██╗ ██╔╝      ██╔══██╗██╔════╝██╔════╝",
-    "  █████╗  ███████║███████╗ ╚████╔╝ █████╗██████╔╝█████╗  █████╗",
-    "  ██╔══╝  ██╔══██║╚════██║  ╚██╔╝  ╚════╝██╔══██╗██╔══╝  ██╔══╝",
-    "  ███████╗██║  ██║███████║   ██║         ██████╔╝███████╗███████╗",
-    "  ╚══════╝╚═╝  ╚═╝╚══════╝   ╚═╝         ╚═════╝ ╚══════╝╚══════╝",
-]
-SUBTITLE = "  Hardware Audit LiveCD"
-
-FG = (0xF6, 0xD0, 0x47)
-FG_DIM = (0xD4, 0xA9, 0x1C)
-SHADOW = (0x5E, 0x47, 0x05)
-SUB = (0x96, 0x7A, 0x17)
-BG = (0x05, 0x05, 0x05)
-
-MONO_FONT_CANDIDATES = [
-    '/usr/share/fonts/truetype/dejavu/DejaVuSansMono-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation2/LiberationMono-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation/LiberationMono-Bold.ttf',
-    '/usr/share/fonts/truetype/freefont/FreeMonoBold.ttf',
-]
-SUB_FONT_CANDIDATES = [
-    '/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation2/LiberationSans-Bold.ttf',
-    '/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf',
-    '/usr/share/fonts/truetype/freefont/FreeSansBold.ttf',
-]
-
-
-def load_font(candidates, size):
-    for path in candidates:
-        if os.path.exists(path):
-            return ImageFont.truetype(path, size)
-    return ImageFont.load_default()
-
-
-def mono_metrics(font):
-    probe = Image.new('L', (W, H), 0)
-    draw = ImageDraw.Draw(probe)
-    char_w = int(round(draw.textlength("M", font=font)))
-    bb = draw.textbbox((0, 0), "Mg", font=font)
-    char_h = bb[3] - bb[1]
-    return char_w, char_h
-
-
-def render_ascii_mask(font, lines, char_w, char_h, line_gap):
-    width = max(len(line) for line in lines) * char_w
-    height = len(lines) * char_h + line_gap * (len(lines) - 1)
-    mask = Image.new('L', (width, height), 0)
-    draw = ImageDraw.Draw(mask)
-    for row, line in enumerate(lines):
-        y = row * (char_h + line_gap)
-        for col, ch in enumerate(line):
-            if ch == ' ':
-                continue
-            x = col * char_w
-            draw.text((x, y), ch, font=font, fill=255)
-    return mask
-
-
-img = Image.new('RGB', (W, H), BG)
-draw = ImageDraw.Draw(img)
-
-# Soft amber glow under the logo without depending on font rendering.
-glow = Image.new('RGBA', (W, H), (0, 0, 0, 0))
-glow_draw = ImageDraw.Draw(glow)
-glow_draw.ellipse((360, 250, 1560, 840), fill=(180, 120, 10, 56))
-glow_draw.ellipse((520, 340, 1400, 760), fill=(255, 190, 40, 36))
-glow = glow.filter(ImageFilter.GaussianBlur(60))
-img = Image.alpha_composite(img.convert('RGBA'), glow)
-
-TARGET_LOGO_W = 400
-max_chars = max(len(line) for line in ASCII_ART)
-_probe_font = load_font(MONO_FONT_CANDIDATES, 64)
-_probe_cw, _ = mono_metrics(_probe_font)
-font_size_logo = max(6, int(64 * TARGET_LOGO_W / (_probe_cw * max_chars)))
-font_logo = load_font(MONO_FONT_CANDIDATES, font_size_logo)
-char_w, char_h = mono_metrics(font_logo)
-logo_mask = render_ascii_mask(font_logo, ASCII_ART, char_w, char_h, 2)
-logo_w, logo_h = logo_mask.size
-logo_x = (W - logo_w) // 2
-logo_y = 380
-
-sh_off = max(1, font_size_logo // 6)
-shadow_mask = logo_mask.filter(ImageFilter.GaussianBlur(1))
-img.paste(SHADOW, (logo_x + sh_off * 2, logo_y + sh_off * 2), shadow_mask)
-img.paste(FG_DIM, (logo_x + sh_off, logo_y + sh_off), logo_mask)
-img.paste(FG, (logo_x, logo_y), logo_mask)
-
-font_sub = load_font(SUB_FONT_CANDIDATES, 30)
-sub_bb = draw.textbbox((0, 0), SUBTITLE, font=font_sub)
-sub_x = (W - (sub_bb[2] - sub_bb[0])) // 2
-sub_y = logo_y + logo_h + 48
-draw = ImageDraw.Draw(img)
-draw.text((sub_x + 2, sub_y + 2), SUBTITLE, font=font_sub, fill=(35, 28, 6))
-draw.text((sub_x, sub_y), SUBTITLE, font=font_sub, fill=SUB)
-
-img = img.convert('RGB')
-
-img.save('/usr/share/bee/wallpaper.png', optimize=True)
-print('wallpaper written: /usr/share/bee/wallpaper.png')
-PYEOF
-
-echo "=== wallpaper done ==="