Make power benchmark report phase-averaged

2026-04-20 10:53:53 +03:00
parent 5f0103635b
commit 6caace0cc0
1 changed files with 285 additions and 71 deletions
--- a/audit/internal/platform/benchmark.go
+++ b/audit/internal/platform/benchmark.go
@@ -67,6 +67,13 @@ type benchmarkPowerCalibrationResult struct {
 	MetricRows []GPUMetricRow
 }

+type benchmarkPowerCalibrationRunSummary struct {
+	LoadedSDR          benchmarkSDRSeriesSummary
+	AvgFanRPM          float64
+	AvgFanDutyCyclePct float64
+	FanSamples         int
+}
+
 type benchmarkBurnProfile struct {
 	name       string
 	category   string
@@ -2413,6 +2420,16 @@ type sdrPowerSnapshot struct {
 	SkippedSensors []string // sensors rejected during self-healing
 }

+type benchmarkSDRSeriesSummary struct {
+	PSUInW   float64
+	PSUOutW  float64
+	GPUSlotW float64
+	PSUSlots map[string]BenchmarkPSUSlotPower
+	Samples  int
+
+	SkippedSensors []string
+}
+
 // sdrSensor is a name+watts pair used for GPU slot self-healing filtering.
 type sdrSensor struct {
 	name  string
@@ -2542,6 +2559,137 @@ func sampleIPMISDRPowerSensors() sdrPowerSnapshot {
 	return snap
 }

+func startIPMISDRSampler(stopCh <-chan struct{}, intervalSec int) <-chan []sdrPowerSnapshot {
+	if intervalSec <= 0 {
+		intervalSec = benchmarkPowerAutotuneSampleInterval
+	}
+	ch := make(chan []sdrPowerSnapshot, 1)
+	go func() {
+		defer close(ch)
+		var samples []sdrPowerSnapshot
+		record := func() {
+			snap := sampleIPMISDRPowerSensors()
+			if snap.PSUInW <= 0 && snap.PSUOutW <= 0 && snap.GPUSlotW <= 0 && len(snap.PSUSlots) == 0 {
+				return
+			}
+			samples = append(samples, snap)
+		}
+		record()
+		ticker := time.NewTicker(time.Duration(intervalSec) * time.Second)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-stopCh:
+				ch <- samples
+				return
+			case <-ticker.C:
+				record()
+			}
+		}
+	}()
+	return ch
+}
+
+func summarizeSDRPowerSeries(samples []sdrPowerSnapshot) benchmarkSDRSeriesSummary {
+	var summary benchmarkSDRSeriesSummary
+	if len(samples) == 0 {
+		return summary
+	}
+
+	type slotAggregate struct {
+		inputs  []float64
+		outputs []float64
+		status  string
+	}
+
+	slotAgg := make(map[string]*slotAggregate)
+	skippedSet := make(map[string]struct{})
+	var inputTotals []float64
+	var outputTotals []float64
+	var gpuSlotTotals []float64
+
+	for _, sample := range samples {
+		if sample.PSUInW > 0 {
+			inputTotals = append(inputTotals, sample.PSUInW)
+		}
+		if sample.PSUOutW > 0 {
+			outputTotals = append(outputTotals, sample.PSUOutW)
+		}
+		if sample.GPUSlotW > 0 {
+			gpuSlotTotals = append(gpuSlotTotals, sample.GPUSlotW)
+		}
+		for _, skipped := range sample.SkippedSensors {
+			if skipped != "" {
+				skippedSet[skipped] = struct{}{}
+			}
+		}
+		for slot, reading := range sample.PSUSlots {
+			agg := slotAgg[slot]
+			if agg == nil {
+				agg = &slotAggregate{}
+				slotAgg[slot] = agg
+			}
+			if reading.InputW != nil && *reading.InputW > 0 {
+				agg.inputs = append(agg.inputs, *reading.InputW)
+			}
+			if reading.OutputW != nil && *reading.OutputW > 0 {
+				agg.outputs = append(agg.outputs, *reading.OutputW)
+			}
+			switch {
+			case reading.Status == "":
+			case agg.status == "":
+				agg.status = reading.Status
+			case agg.status == "OK" && reading.Status != "OK":
+				agg.status = reading.Status
+			}
+		}
+	}
+
+	summary.PSUInW = benchmarkMean(inputTotals)
+	summary.PSUOutW = benchmarkMean(outputTotals)
+	summary.GPUSlotW = benchmarkMean(gpuSlotTotals)
+	summary.Samples = len(samples)
+
+	if len(slotAgg) > 0 {
+		summary.PSUSlots = make(map[string]BenchmarkPSUSlotPower, len(slotAgg))
+		for slot, agg := range slotAgg {
+			reading := BenchmarkPSUSlotPower{Status: agg.status}
+			if mean := benchmarkMean(agg.inputs); mean > 0 {
+				v := mean
+				reading.InputW = &v
+			}
+			if mean := benchmarkMean(agg.outputs); mean > 0 {
+				v := mean
+				reading.OutputW = &v
+			}
+			summary.PSUSlots[slot] = reading
+		}
+	}
+	if len(skippedSet) > 0 {
+		summary.SkippedSensors = make([]string, 0, len(skippedSet))
+		for skipped := range skippedSet {
+			summary.SkippedSensors = append(summary.SkippedSensors, skipped)
+		}
+		sort.Strings(summary.SkippedSensors)
+	}
+
+	return summary
+}
+
+func collectIPMISDRPowerSeries(ctx context.Context, durationSec, intervalSec int) benchmarkSDRSeriesSummary {
+	if durationSec <= 0 {
+		return benchmarkSDRSeriesSummary{}
+	}
+	stopCh := make(chan struct{})
+	doneCh := startIPMISDRSampler(stopCh, intervalSec)
+	select {
+	case <-ctx.Done():
+	case <-time.After(time.Duration(durationSec) * time.Second):
+	}
+	close(stopCh)
+	return summarizeSDRPowerSeries(<-doneCh)
+}
+
 // queryIPMIServerPowerW reads the current server power draw via ipmitool dcmi.
 // Returns 0 and an error if IPMI is unavailable or the output cannot be parsed.
 func queryIPMIServerPowerW() (float64, error) {
@@ -3086,8 +3234,9 @@ func runBenchmarkPowerCalibration(
 	logFunc func(string),
 	seedLimits map[int]int,
 	durationSec int,
-) (map[int]benchmarkPowerCalibrationResult, []benchmarkRestoreAction, []GPUMetricRow) {
+) (map[int]benchmarkPowerCalibrationResult, []benchmarkRestoreAction, []GPUMetricRow, benchmarkPowerCalibrationRunSummary) {
 	calibDurationSec := durationSec
+	var runSummary benchmarkPowerCalibrationRunSummary
 	if calibDurationSec <= 0 {
 		calibDurationSec = 120
 	}
@@ -3105,12 +3254,12 @@ func runBenchmarkPowerCalibration(
 	if engine == BenchmarkPowerEngineTargetedPower {
 		if _, err := exec.LookPath("dcgmi"); err != nil {
 			logFunc("power calibration: dcgmi not found, skipping (will use default power limit)")
-			return map[int]benchmarkPowerCalibrationResult{}, nil, nil
+			return map[int]benchmarkPowerCalibrationResult{}, nil, nil, runSummary
 		}
 	} else {
 		if _, _, err := resolveBenchmarkPowerLoadCommand(calibDurationSec, gpuIndices); err != nil {
 			logFunc("power calibration: dcgmproftester not found, skipping (will use default power limit)")
-			return map[int]benchmarkPowerCalibrationResult{}, nil, nil
+			return map[int]benchmarkPowerCalibrationResult{}, nil, nil, runSummary
 		}
 	}
 	if killed := KillTestWorkers(); len(killed) > 0 {
@@ -3275,6 +3424,10 @@ calibDone:
 		}
 		attemptCtx, cancelAttempt := context.WithCancel(ctx)
 		doneCh := make(chan sharedAttemptResult, 1)
+		sdrStopCh := make(chan struct{})
+		sdrDoneCh := startIPMISDRSampler(sdrStopCh, benchmarkPowerAutotuneSampleInterval)
+		fanStopCh := make(chan struct{})
+		fanDoneCh := startBenchmarkFanSampler(fanStopCh, benchmarkPowerAutotuneSampleInterval)
 		go func() {
 			out, rows, err := runBenchmarkCommandWithMetrics(attemptCtx, verboseLog, logName, cmd, env, gpuIndices, logFunc)
 			doneCh <- sharedAttemptResult{out: out, rows: rows, err: err}
@@ -3314,6 +3467,10 @@ calibDone:
 		}
 		ticker.Stop()
 		cancelAttempt()
+		close(sdrStopCh)
+		close(fanStopCh)
+		attemptSDRSummary := summarizeSDRPowerSeries(<-sdrDoneCh)
+		attemptFanSummary := <-fanDoneCh
 		_ = os.WriteFile(filepath.Join(runDir, logName), ar.out, 0644)
 		// Accumulate telemetry rows with attempt stage label.
 		appendBenchmarkMetrics(&allCalibRows, ar.rows, fmt.Sprintf("attempt-%d", sharedAttempt), &calibCursor, float64(calibDurationSec))
@@ -3351,10 +3508,14 @@ calibDone:
 		busyDelaySec = 1

 		// Per-GPU analysis and binary search update.
+		attemptStable := ar.err == nil
 		for _, s := range active {
 			perGPU := filterRowsByGPU(ar.rows, s.idx)
 			summary := summarizeBenchmarkTelemetry(perGPU)
 			throttle := throttleReasons[s.idx]
+			if throttle != "" || summary.P95PowerW <= 0 {
+				attemptStable = false
+			}

 			// Cooling warning: thermal throttle with fans not at maximum.
 			if strings.Contains(throttle, "thermal") && s.calib.CoolingWarning == "" {
@@ -3487,6 +3648,16 @@ calibDone:
 			s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: trying %d W (lo=%d hi=%d)", next, s.lo, s.hi))
 			logFunc(fmt.Sprintf("power calibration: GPU %d binary search: trying %d W (lo=%d hi=%d)", s.idx, next, s.lo, s.hi))
 		}
+		if attemptStable {
+			if attemptSDRSummary.Samples > 0 {
+				runSummary.LoadedSDR = attemptSDRSummary
+			}
+			if attemptFanSummary.FanSamples > 0 {
+				runSummary.AvgFanRPM = attemptFanSummary.AvgFanRPM
+				runSummary.AvgFanDutyCyclePct = attemptFanSummary.AvgFanDutyCyclePct
+				runSummary.FanSamples = attemptFanSummary.FanSamples
+			}
+		}
 	}

 	for _, s := range states {
@@ -3495,7 +3666,7 @@ calibDone:
 		}
 	}
 	writeBenchmarkMetricsFiles(runDir, allCalibRows)
-	return results, restore, allCalibRows
+	return results, restore, allCalibRows, runSummary
 }

 // isDCGMResourceBusy returns true when dcgmi exits with DCGM_ST_IN_USE (222),
@@ -3540,6 +3711,47 @@ func meanFanRPM(fans []FanReading) float64 {
 	return sum / float64(len(fans))
 }

+func startBenchmarkFanSampler(stopCh <-chan struct{}, intervalSec int) <-chan benchmarkPowerCalibrationRunSummary {
+	if intervalSec <= 0 {
+		intervalSec = benchmarkPowerAutotuneSampleInterval
+	}
+	ch := make(chan benchmarkPowerCalibrationRunSummary, 1)
+	go func() {
+		defer close(ch)
+		var rpmSamples []float64
+		var dutySamples []float64
+		record := func() {
+			fans, err := sampleFanSpeeds()
+			if err != nil || len(fans) == 0 {
+				return
+			}
+			if rpm := meanFanRPM(fans); rpm > 0 {
+				rpmSamples = append(rpmSamples, rpm)
+			}
+			if duty, ok, _ := sampleFanDutyCyclePctFromFans(fans); ok && duty > 0 {
+				dutySamples = append(dutySamples, duty)
+			}
+		}
+		record()
+		ticker := time.NewTicker(time.Duration(intervalSec) * time.Second)
+		defer ticker.Stop()
+		for {
+			select {
+			case <-stopCh:
+				ch <- benchmarkPowerCalibrationRunSummary{
+					AvgFanRPM:          benchmarkMean(rpmSamples),
+					AvgFanDutyCyclePct: benchmarkMean(dutySamples),
+					FanSamples:         len(rpmSamples),
+				}
+				return
+			case <-ticker.C:
+				record()
+			}
+		}
+	}()
+	return ch
+}
+
 func powerBenchDurationSec(profile string) int {
 	switch strings.TrimSpace(strings.ToLower(profile)) {
 	case NvidiaBenchmarkProfileStability:
@@ -3568,41 +3780,39 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 	fmt.Fprintf(&b, "**Overall status:** %s  \n", result.OverallStatus)
 	fmt.Fprintf(&b, "**Platform max TDP (GPU-reported):** %.0f W  \n", result.PlatformMaxTDPW)
 	if sp := result.ServerPower; sp != nil && sp.Available {
-		fmt.Fprintf(&b, "**Server power delta (IPMI DCMI):** %.0f W  \n", sp.DeltaW)
-		if sp.PSUInputLoadedW > 0 {
-			psuDelta := sp.PSUInputLoadedW - sp.PSUInputIdleW
-			fmt.Fprintf(&b, "**PSU AC input Δ (IPMI SDR):** %.0f W  \n", psuDelta)
+		sourceLabel := "autotuned source"
+		switch normalizeBenchmarkPowerSource(sp.Source) {
+		case BenchmarkPowerSourceSDRPSUInput:
+			sourceLabel = "autotuned source (SDR PSU AC input)"
+		case BenchmarkPowerSourceDCMI:
+			sourceLabel = "autotuned source (DCMI)"
 		}
-		fmt.Fprintf(&b, "**Reporting ratio (IPMI Δ / GPU actual sum):** %.2f  \n", sp.ReportingRatio)
+		fmt.Fprintf(&b, "**Server power delta (%s):** %.0f W  \n", sourceLabel, sp.DeltaW)
+		fmt.Fprintf(&b, "**Reporting ratio:** %.2f  \n", sp.ReportingRatio)
 	}
 	b.WriteString("\n")
 	// Server power comparison table.
 	if sp := result.ServerPower; sp != nil {
 		b.WriteString("## Server vs GPU Power Comparison\n\n")
+		selectedSource := normalizeBenchmarkPowerSource(sp.Source)
+		selectedSourceLabel := "Selected source"
+		if selectedSource == BenchmarkPowerSourceSDRPSUInput {
+			selectedSourceLabel = "Selected source (SDR PSU AC input)"
+		} else if selectedSource == BenchmarkPowerSourceDCMI {
+			selectedSourceLabel = "Selected source (DCMI)"
+		}
 		var spRows [][]string
-		spRows = append(spRows, []string{"GPU stable limits sum", "nvidia-smi", fmt.Sprintf("%.0f W", result.PlatformMaxTDPW)})
-		spRows = append(spRows, []string{"GPU actual power sum (p95, last step)", "nvidia-smi", fmt.Sprintf("%.0f W", sp.GPUReportedSumW)})
-		if sp.GPUSlotTotalW > 0 {
-			spRows = append(spRows, []string{"GPU PCIe slot power (at peak load)", "IPMI SDR", fmt.Sprintf("%.0f W", sp.GPUSlotTotalW)})
-		}
+		spRows = append(spRows, []string{"GPU actual power sum (p95, last step)", fmt.Sprintf("%.0f W", sp.GPUReportedSumW)})
 		if sp.Available {
-			spRows = append(spRows, []string{"Server idle power", "IPMI DCMI", fmt.Sprintf("%.0f W", sp.IdleW)})
-			spRows = append(spRows, []string{"Server loaded power", "IPMI DCMI", fmt.Sprintf("%.0f W", sp.LoadedW)})
-			spRows = append(spRows, []string{"Server Δ power (loaded − idle)", "IPMI DCMI", fmt.Sprintf("%.0f W", sp.DeltaW)})
+			spRows = append(spRows, []string{selectedSourceLabel + " idle power", fmt.Sprintf("%.0f W", sp.IdleW)})
+			spRows = append(spRows, []string{selectedSourceLabel + " loaded power", fmt.Sprintf("%.0f W", sp.LoadedW)})
+			spRows = append(spRows, []string{selectedSourceLabel + " Δ power (loaded − idle)", fmt.Sprintf("%.0f W", sp.DeltaW)})
 		}
-		if sp.PSUInputLoadedW > 0 {
-			spRows = append(spRows, []string{"PSU AC input (idle)", "IPMI SDR", fmt.Sprintf("%.0f W", sp.PSUInputIdleW)})
-			spRows = append(spRows, []string{"PSU AC input (loaded)", "IPMI SDR", fmt.Sprintf("%.0f W", sp.PSUInputLoadedW)})
+		if selectedSource == BenchmarkPowerSourceSDRPSUInput && sp.PSUInputLoadedW > 0 {
+			spRows = append(spRows, []string{"PSU AC input (idle avg, pre-load phase)", fmt.Sprintf("%.0f W", sp.PSUInputIdleW)})
+			spRows = append(spRows, []string{"PSU AC input (loaded avg, final phase)", fmt.Sprintf("%.0f W", sp.PSUInputLoadedW)})
 			psuDelta := sp.PSUInputLoadedW - sp.PSUInputIdleW
-			spRows = append(spRows, []string{"PSU AC input Δ (loaded − idle)", "IPMI SDR", fmt.Sprintf("%.0f W", psuDelta)})
-		}
-		if sp.PSUOutputLoadedW > 0 {
-			spRows = append(spRows, []string{"PSU DC output (idle)", "IPMI SDR", fmt.Sprintf("%.0f W", sp.PSUOutputIdleW)})
-			spRows = append(spRows, []string{"PSU DC output (loaded)", "IPMI SDR", fmt.Sprintf("%.0f W", sp.PSUOutputLoadedW)})
-			if sp.PSUInputLoadedW > 0 && sp.PSUInputIdleW > 0 {
-				psuEff := sp.PSUOutputIdleW / sp.PSUInputIdleW * 100
-				spRows = append(spRows, []string{"PSU conversion efficiency (idle)", "IPMI SDR", fmt.Sprintf("%.1f%%", psuEff)})
-			}
+			spRows = append(spRows, []string{"PSU AC input Δ (loaded − idle)", fmt.Sprintf("%.0f W", psuDelta)})
 		}
 		if sp.Available {
 			ratio := sp.ReportingRatio
@@ -3619,8 +3829,8 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 			default:
 				ratioNote = "✗ significant discrepancy — GPU over-reports TDP vs wall power"
 			}
-			spRows = append(spRows, []string{"Reporting ratio (DCMI Δ / GPU actual)", "IPMI DCMI", fmt.Sprintf("%.2f — %s", ratio, ratioNote)})
-			if sp.PSUInputLoadedW > 0 && sp.GPUReportedSumW > 0 {
+			spRows = append(spRows, []string{"Reporting ratio", fmt.Sprintf("%.2f — %s", ratio, ratioNote)})
+			if selectedSource == BenchmarkPowerSourceSDRPSUInput && sp.PSUInputLoadedW > 0 && sp.GPUReportedSumW > 0 {
 				psuDelta := sp.PSUInputLoadedW - sp.PSUInputIdleW
 				sdrRatio := psuDelta / sp.GPUReportedSumW
 				sdrNote := ""
@@ -3632,12 +3842,12 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				default:
 					sdrNote = "✗ significant discrepancy"
 				}
-				spRows = append(spRows, []string{"Reporting ratio (SDR PSU Δ / GPU actual)", "IPMI SDR", fmt.Sprintf("%.2f — %s", sdrRatio, sdrNote)})
+				spRows = append(spRows, []string{"PSU AC input reporting ratio", fmt.Sprintf("%.2f — %s", sdrRatio, sdrNote)})
 			}
 		} else {
-			spRows = append(spRows, []string{"IPMI availability", "—", "not available — IPMI not supported or ipmitool not found"})
+			spRows = append(spRows, []string{"IPMI availability", "not available — IPMI not supported or ipmitool not found"})
 		}
-		b.WriteString(fmtMDTable([]string{"Metric", "Source", "Value"}, spRows))
+		b.WriteString(fmtMDTable([]string{"Metric", "Value"}, spRows))
 		for _, note := range sp.Notes {
 			fmt.Fprintf(&b, "\n> %s\n", note)
 		}
@@ -3689,11 +3899,10 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				psuDistRows = append(psuDistRows, []string{
 					slot,
 					fmtW(idle.InputW), fmtW(loaded.InputW),
-					fmtW(idle.OutputW), fmtW(loaded.OutputW),
 					deltaStr, status,
 				})
 			}
-			b.WriteString(fmtMDTable([]string{"Slot", "AC Input (idle)", "AC Input (loaded)", "DC Output (idle)", "DC Output (loaded)", "Load Δ", "Status"}, psuDistRows))
+			b.WriteString(fmtMDTable([]string{"Slot", "AC Input (idle avg)", "AC Input (loaded avg)", "Load Δ", "Status"}, psuDistRows))
 			b.WriteString("\n")
 		}
 	}
@@ -3741,7 +3950,7 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				fan,
 			})
 		}
-		b.WriteString(fmtMDTable([]string{"GPU", "Clock MHz (Mem MHz)", "Avg Temp °C", "Power W", "Server Δ W", "Fan RPM (duty%)"}, sgRows))
+		b.WriteString(fmtMDTable([]string{"GPU", "Clock MHz (Mem MHz)", "Avg Temp °C", "Power W", "Server Δ W", "Avg Fan RPM (duty%)"}, sgRows))
 		b.WriteString("\n")
 	}
 	if len(result.RecommendedSlotOrder) > 0 {
@@ -3850,7 +4059,7 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 			for _, slot := range psuSlots {
 				psuHeaders = append(psuHeaders, fmt.Sprintf("PSU %s W", slot))
 			}
-			psuHeaders = append(psuHeaders, "PSU Total W", "Platform eff.", "Fan RPM (duty%)")
+			psuHeaders = append(psuHeaders, "PSU Total W", "Platform eff.", "Avg Fan RPM (duty%)")

 			var psuRows [][]string
 			for _, step := range result.RampSteps {
@@ -3931,7 +4140,6 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 			}
 			pdRows = append(pdRows, []string{
 				fmt.Sprintf("GPU %d", gpu.Index),
-				fmt.Sprintf("%.0f W", gpu.DefaultPowerLimitW),
 				fmt.Sprintf("%.0f W", gpu.AppliedPowerLimitW),
 				fmt.Sprintf("%.0f W", stable),
 				realization,
@@ -3944,13 +4152,12 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 		}
 		pdRows = append(pdRows, []string{
 			"**Platform**",
-			fmt.Sprintf("**%.0f W**", totalDefault),
 			"—",
 			fmt.Sprintf("**%.0f W**", totalStable),
 			fmt.Sprintf("**%s**", platformReal),
 			"",
 		})
-		b.WriteString(fmtMDTable([]string{"GPU", "Default TDP", "Single-card limit", "Stable limit", "Realization", "Derated"}, pdRows))
+		b.WriteString(fmtMDTable([]string{"GPU", "Single-card limit", "Stable limit", "Realization", "Derated"}, pdRows))
 		b.WriteString("\n")

 		// Balance across GPUs — only meaningful with 2+ GPUs.
@@ -4100,7 +4307,7 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 			{"Avg Temp °C", singleTemp},
 			{"Power W", singlePwr},
 			{"Per GPU wall W", singleWall},
-			{"Fan RPM (duty%)", singleFan},
+			{"Avg Fan RPM (duty%)", singleFan},
 		}
 		if lastStep != nil {
 			compRows[0] = append(compRows[0], fmt.Sprintf("%s (%s)", allClk, allMem))
@@ -4208,18 +4415,22 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	// Sample server idle power before any GPU load.
 	var serverIdleW float64
 	var serverIdleOK bool
+	idleSDRStopCh := make(chan struct{})
+	idleSDRCh := startIPMISDRSampler(idleSDRStopCh, benchmarkPowerAutotuneSampleInterval)
 	if w, ok := sampleBenchmarkPowerSourceSeries(ctx, opts.ServerPowerSource, 10, benchmarkPowerAutotuneSampleInterval); ok {
 		serverIdleW = w
 		serverIdleOK = true
 		logFunc(fmt.Sprintf("server idle power (%s): %.0f W", opts.ServerPowerSource, w))
 	}
-	sdrIdle := sampleIPMISDRPowerSensors()
+	close(idleSDRStopCh)
+	sdrIdle := summarizeSDRPowerSeries(<-idleSDRCh)
 	psuBefore := psuStatusSnapshot()

 	// Phase 1: calibrate each GPU individually (sequentially, one at a time) to
 	// establish a true single-card power baseline unaffected by neighbour heat.
 	calibByIndex := make(map[int]benchmarkPowerCalibrationResult, len(selected))
 	singleIPMILoadedW := make(map[int]float64, len(selected))
+	singleRunSummaryByIndex := make(map[int]benchmarkPowerCalibrationRunSummary, len(selected))
 	var allRestoreActions []benchmarkRestoreAction
 	// allPowerRows accumulates telemetry from all phases for the top-level gpu-metrics.csv.
 	var allPowerRows []GPUMetricRow
@@ -4235,21 +4446,21 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		logFunc(fmt.Sprintf("power calibration: GPU %d single-card baseline", idx))
 		singlePowerStopCh := make(chan struct{})
 		singlePowerCh := startSelectedPowerSourceSampler(singlePowerStopCh, opts.ServerPowerSource, benchmarkPowerAutotuneSampleInterval)
-		c, restore, singleRows := runBenchmarkPowerCalibration(ctx, verboseLog, singleDir, []int{idx}, singleInfo, logFunc, nil, durationSec)
+		c, restore, singleRows, singleRun := runBenchmarkPowerCalibration(ctx, verboseLog, singleDir, []int{idx}, singleInfo, logFunc, nil, durationSec)
 		appendBenchmarkMetrics(&allPowerRows, singleRows, fmt.Sprintf("single-gpu-%d", idx), &powerCursor, 0)
 		close(singlePowerStopCh)
-		sdrSingle := sampleIPMISDRPowerSensors()
 		if samples := <-singlePowerCh; len(samples) > 0 {
 			singleIPMILoadedW[idx] = benchmarkMean(samples)
 			logFunc(fmt.Sprintf("power calibration: GPU %d single-card server power (%s avg): %.0f W", idx, opts.ServerPowerSource, singleIPMILoadedW[idx]))
-		} else if opts.ServerPowerSource == BenchmarkPowerSourceSDRPSUInput && sdrSingle.PSUInW > 0 {
-			singleIPMILoadedW[idx] = sdrSingle.PSUInW
-			logFunc(fmt.Sprintf("power calibration: GPU %d single-card fallback server power (SDR snapshot): %.0f W", idx, sdrSingle.PSUInW))
+		} else if opts.ServerPowerSource == BenchmarkPowerSourceSDRPSUInput && singleRun.LoadedSDR.PSUInW > 0 {
+			singleIPMILoadedW[idx] = singleRun.LoadedSDR.PSUInW
+			logFunc(fmt.Sprintf("power calibration: GPU %d single-card fallback server power (SDR avg): %.0f W", idx, singleRun.LoadedSDR.PSUInW))
 		}
 		allRestoreActions = append(allRestoreActions, restore...)
 		if r, ok := c[idx]; ok {
 			calibByIndex[idx] = r
 		}
+		singleRunSummaryByIndex[idx] = singleRun
 	}
 	defer func() {
 		for i := len(allRestoreActions) - 1; i >= 0; i-- {
@@ -4292,11 +4503,9 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			t := summarizeBenchmarkTelemetry(calib.MetricRows)
 			gpu.Telemetry = &t
 		}
-		if fans, err := sampleFanSpeeds(); err == nil && len(fans) > 0 {
-			gpu.AvgFanRPM = meanFanRPM(fans)
-			if duty, ok, _ := sampleFanDutyCyclePctFromFans(fans); ok {
-				gpu.AvgFanDutyCyclePct = duty
-			}
+		if singleRun := singleRunSummaryByIndex[idx]; singleRun.AvgFanRPM > 0 {
+			gpu.AvgFanRPM = singleRun.AvgFanRPM
+			gpu.AvgFanDutyCyclePct = singleRun.AvgFanDutyCyclePct
 		}
 		gpus = append(gpus, gpu)
 	}
@@ -4352,10 +4561,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	// per-step in NvidiaPowerBenchStep.ServerLoadedW.
 	var serverLoadedW float64
 	var serverLoadedOK bool
-	// sdrLastStep retains the SDR snapshot from the last ramp step while GPUs are
-	// still loaded. Used as PSUInputLoadedW in the summary instead of re-sampling
-	// after the test when GPUs have already returned to idle.
-	var sdrLastStep sdrPowerSnapshot
+	// sdrLastStep retains the phase-averaged SDR readings from the last ramp step
+	// while GPUs are loaded. Used in the summary instead of re-sampling after the
+	// test when GPUs have already returned to idle.
+	var sdrLastStep benchmarkSDRSeriesSummary

 	// Step 1: reuse single-card calibration result directly.
 	if len(result.RecommendedSlotOrder) > 0 {
@@ -4376,6 +4585,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			ramp.ServerLoadedW = w
 			ramp.ServerDeltaW = w - serverIdleW
 		}
+		if singleRun := singleRunSummaryByIndex[firstIdx]; singleRun.AvgFanRPM > 0 {
+			ramp.AvgFanRPM = singleRun.AvgFanRPM
+			ramp.AvgFanDutyCyclePct = singleRun.AvgFanDutyCyclePct
+		}
 		if !firstCalib.Completed {
 			ramp.Status = "FAILED"
 			ramp.Notes = append(ramp.Notes, fmt.Sprintf("GPU %d did not complete single-card %s", firstIdx, benchmarkPowerEngineLabel(benchmarkPowerEngine())))
@@ -4426,7 +4639,7 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		stepInfo := cloneBenchmarkGPUInfoMap(infoByIndex)
 		stepPowerStopCh := make(chan struct{})
 		stepPowerCh := startSelectedPowerSourceSampler(stepPowerStopCh, opts.ServerPowerSource, benchmarkPowerAutotuneSampleInterval)
-		stepCalib, stepRestore, stepRows := runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc, seedForStep, durationSec)
+		stepCalib, stepRestore, stepRows, stepRun := runBenchmarkPowerCalibration(ctx, verboseLog, stepDir, subset, stepInfo, logFunc, seedForStep, durationSec)
 		appendBenchmarkMetrics(&allPowerRows, stepRows, fmt.Sprintf("ramp-step-%d", step), &powerCursor, 0)
 		close(stepPowerStopCh)
 		var stepIPMILoadedW float64
@@ -4497,10 +4710,9 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			result.Findings = append(result.Findings, fmt.Sprintf("Ramp step %d (GPU %d) required derating to %.0f W under combined thermal load.", step, newGPUIdx, c.AppliedPowerLimitW))
 		}

-		// Per-step PSU slot snapshot — also used as the authoritative loaded power
-		// source when SDR PSU sensors are available (more accurate than DCMI on
-		// servers where DCMI covers only a subset of installed PSUs).
-		sdrStep := sampleIPMISDRPowerSensors()
+		// Per-step PSU slot readings are averaged over the whole load phase rather
+		// than captured as a single end-of-phase snapshot.
+		sdrStep := stepRun.LoadedSDR
 		if len(sdrStep.PSUSlots) > 0 {
 			ramp.PSUSlotReadings = sdrStep.PSUSlots
 		}
@@ -4518,7 +4730,7 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 		} else if opts.ServerPowerSource == BenchmarkPowerSourceSDRPSUInput && sdrStep.PSUInW > 0 {
 			ramp.ServerLoadedW = sdrStep.PSUInW
 			ramp.ServerDeltaW = sdrStep.PSUInW - sdrIdle.PSUInW
-			logFunc(fmt.Sprintf("power ramp: step %d fallback server loaded power (SDR snapshot): %.0f W", step, sdrStep.PSUInW))
+			logFunc(fmt.Sprintf("power ramp: step %d fallback server loaded power (SDR avg): %.0f W", step, sdrStep.PSUInW))
 			if step == len(result.RecommendedSlotOrder) {
 				serverLoadedW = sdrStep.PSUInW
 				serverLoadedOK = true
@@ -4526,12 +4738,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			}
 		}

-		// Fan state at end of ramp step.
-		if fans, err := sampleFanSpeeds(); err == nil && len(fans) > 0 {
-			ramp.AvgFanRPM = meanFanRPM(fans)
-			if duty, ok, _ := sampleFanDutyCyclePctFromFans(fans); ok {
-				ramp.AvgFanDutyCyclePct = duty
-			}
+		// Fan values are phase averages over the same load window.
+		if stepRun.AvgFanRPM > 0 {
+			ramp.AvgFanRPM = stepRun.AvgFanRPM
+			ramp.AvgFanDutyCyclePct = stepRun.AvgFanDutyCyclePct
 		}

 		// Per-GPU telemetry from this ramp step's calibration.
@@ -4584,8 +4794,8 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 	// Supplement DCMI with SDR multi-source data via collector's PSU slot patterns.
 	// Per-slot readings enable correlation with audit HardwarePowerSupply entries.
 	if result.ServerPower != nil {
-		// Use the SDR snapshot from the last ramp step (GPUs still loaded) rather
-		// than re-sampling here, which would capture post-test idle state.
+		// Use the SDR phase average from the last ramp step (GPUs still loaded)
+		// rather than re-sampling here, which would capture post-test idle state.
 		sdrLoaded := sdrLastStep
 		result.ServerPower.PSUInputIdleW = sdrIdle.PSUInW
 		result.ServerPower.PSUInputLoadedW = sdrLoaded.PSUInW
@@ -4605,6 +4815,10 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			result.ServerPower.Notes = append(result.ServerPower.Notes,
 				"SDR sensors skipped (self-healed): "+strings.Join(sdrLoaded.SkippedSensors, "; "))
 		}
+		if sdrLoaded.Samples > 0 {
+			result.ServerPower.Notes = append(result.ServerPower.Notes,
+				fmt.Sprintf("Final SDR PSU loaded values are phase averages across %d sample(s) from the last full-load step.", sdrLoaded.Samples))
+		}
 		// Detect DCMI partial coverage: direct SDR comparison first,
 		// ramp heuristic as fallback when SDR PSU sensors are absent.
 		dcmiUnreliable := detectDCMIPartialCoverage(result.ServerPower) ||