Run power calibration for all GPUs simultaneously

Previously each GPU was calibrated sequentially (one card fully done
before the next started), producing the staircase temperature pattern
seen on the graph.

Now all GPUs run together in a single dcgmi diag -r targeted_power
session per attempt. This means:
- All cards are under realistic thermal load at the same time.
- A single DCGM session handles the run — no resource-busy contention
  from concurrent dcgmi processes.
- Binary search state (lo/hi) is tracked independently per GPU; each
  card converges to its own highest stable power limit.
- Throttle counter polling covers all active GPUs in the shared ticker.
- Resource-busy exponential back-off is shared (one DCGM session).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/platform/benchmark.go

@@ -2476,9 +2476,6 @@ func runBenchmarkPowerCalibration(
 	// calibSearchTolerance is the binary-search convergence threshold in watts.
 	// When hi-lo ≤ this, the highest verified-stable limit (lo) is used.
 	const calibSearchTolerance = 10
-	// calibPreThrottleMarginW is subtracted from the telemetry-estimated
-	// pre-throttle power draw to produce a smarter initial search candidate.
-	const calibPreThrottleMarginW = 10
 	// dcgmResourceBusyMaxDelaySec caps the exponential back-off when DCGM
 	// returns DCGM_ST_IN_USE (exit 222). The sequence is 1 s, 2 s, 4 s, …
 	// doubling each retry until it would exceed the cap, at which point the
@@ -2501,8 +2498,25 @@ func runBenchmarkPowerCalibration(
 		err  error
 	}
 
+
+	// gpuCalibState holds per-GPU binary search state during parallel calibration.
+	type gpuCalibState struct {
+		idx            int
+		info           benchmarkGPUInfo
+		originalLimitW int
+		appliedLimitW  int
+		minLimitW      int
+		lo             int // highest verified-stable limit (assumed: minLimitW)
+		hi             int // lowest verified-unstable limit (exclusive sentinel above start)
+		calib          benchmarkPowerCalibrationResult
+		converged      bool
+	}
+
 	results := make(map[int]benchmarkPowerCalibrationResult, len(gpuIndices))
 	var restore []benchmarkRestoreAction
+
+	// Initialise per-GPU state.
+	states := make([]*gpuCalibState, 0, len(gpuIndices))
 	for _, idx := range gpuIndices {
 		info := infoByIndex[idx]
 		originalLimitW := int(math.Round(info.PowerLimitW))
@@ -2531,17 +2545,17 @@ func runBenchmarkPowerCalibration(
 		if minLimitW < calibSearchTolerance {
 			minLimitW = calibSearchTolerance
 		}
-
+		s := &gpuCalibState{
-		calib := benchmarkPowerCalibrationResult{
+			idx:            idx,
-			AppliedPowerLimitW: float64(appliedLimitW),
+			info:           info,
+			originalLimitW: originalLimitW,
+			appliedLimitW:  appliedLimitW,
+			minLimitW:      minLimitW,
+			lo:             minLimitW,
+			hi:             appliedLimitW + 1, // not yet tested, not yet confirmed unstable
+			calib:          benchmarkPowerCalibrationResult{AppliedPowerLimitW: float64(appliedLimitW)},
 		}
-		// Binary search bounds for finding the highest stable power limit.
+		states = append(states, s)
-		// lo = highest verified-stable level (assumed: minLimitW).
-		// hi = lowest verified-unstable level (assumed: above the starting limit).
-		lo := minLimitW
-		hi := appliedLimitW + 1 // exclusive: not yet tested, so not yet confirmed unstable
-		busyRetries := 0
-		busyDelaySec := 1 // exponential back-off seed; doubles each retry up to dcgmResourceBusyMaxDelaySec
 		if canDerate && originalLimitW > 0 {
 			idxCopy := idx
 			orig := originalLimitW
@@ -2552,212 +2566,243 @@ func runBenchmarkPowerCalibration(
 				},
 			})
 		}
+	}
 
-	calibLoop:
+	// Shared DCGM resource-busy back-off state (single diagnostic session).
+	busyRetries := 0
+	busyDelaySec := 1
+	sharedAttempt := 0
+
+	type sharedAttemptResult struct {
+		out  []byte
+		rows []GPUMetricRow
+		err  error
+	}
+
+calibDone:
+	for {
+		// Collect non-converged GPUs.
+		var active []*gpuCalibState
+		for _, s := range states {
+			if !s.converged {
+				active = append(active, s)
+			}
+		}
+		if len(active) == 0 || ctx.Err() != nil {
+			break
+		}
+
+		sharedAttempt++
+		for _, s := range active {
+			s.calib.Attempts++
+			logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power attempt %d at %d W for %ds", s.idx, s.calib.Attempts, s.appliedLimitW, calibDurationSec))
+		}
+
+		// Snapshot throttle counters for all active GPUs before the run.
+		beforeThrottle := make(map[int]BenchmarkThrottleCounters, len(active))
+		for _, s := range active {
+			beforeThrottle[s.idx], _ = queryThrottleCounters(s.idx)
+		}
+
+		// Run targeted_power for ALL gpuIndices simultaneously so every card
+		// is under load during calibration — this reflects real server thermals.
+		logName := fmt.Sprintf("power-calibration-attempt-%d.log", sharedAttempt)
+		cmd := nvidiaDCGMNamedDiagCommand("targeted_power", calibDurationSec, gpuIndices)
+		attemptCtx, cancelAttempt := context.WithCancel(ctx)
+		doneCh := make(chan sharedAttemptResult, 1)
+		go func() {
+			out, rows, err := runBenchmarkCommandWithMetrics(attemptCtx, verboseLog, logName, cmd, nil, gpuIndices, logFunc)
+			doneCh <- sharedAttemptResult{out: out, rows: rows, err: err}
+		}()
+
+		ticker := time.NewTicker(time.Second)
+		throttleReasons := make(map[int]string, len(active))
+		var ar sharedAttemptResult
+
+	attemptLoop:
 		for {
-			calib.Attempts++
+			select {
-			logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power attempt %d at %d W for %ds", idx, calib.Attempts, appliedLimitW, calibDurationSec))
+			case ar = <-doneCh:
-
+				break attemptLoop
-			beforeThrottle, _ := queryThrottleCounters(idx)
+			case <-ticker.C:
-			attemptCtx, cancel := context.WithCancel(ctx)
+				// Poll throttle counters for each active GPU independently.
-			doneCh := make(chan calibrationAttemptResult, 1)
+				for _, s := range active {
-			logName := fmt.Sprintf("power-calibration-gpu-%d-attempt-%d.log", idx, calib.Attempts)
+					if throttleReasons[s.idx] != "" {
-			cmd := nvidiaDCGMNamedDiagCommand("targeted_power", calibDurationSec, []int{idx})
+						continue // already detected for this GPU
-			go func() {
+					}
-				out, rows, err := runBenchmarkCommandWithMetrics(attemptCtx, verboseLog, logName, cmd, nil, []int{idx}, logFunc)
+					after, err := queryThrottleCounters(s.idx)
-				doneCh <- calibrationAttemptResult{out: out, rows: rows, err: err}
-			}()
-
-			ticker := time.NewTicker(time.Second)
-			var (
-				attempt        calibrationAttemptResult
-				throttleReason string
-			)
-		attemptLoop:
-			for {
-				select {
-				case attempt = <-doneCh:
-					break attemptLoop
-				case <-ticker.C:
-					afterThrottle, err := queryThrottleCounters(idx)
 					if err != nil {
 						continue
 					}
-					// Record the throttle reason but do NOT cancel the dcgmi
+					// Record throttle but do NOT cancel — let dcgmi finish so
-					// process. Killing it mid-run leaves nv-hostengine holding
+					// nv-hostengine releases the slot cleanly before the next attempt.
-					// the diagnostic slot, which causes DCGM_ST_IN_USE on every
+					if reason := benchmarkCalibrationThrottleReason(beforeThrottle[s.idx], after); reason != "" {
-					// subsequent attempt. Let targeted_power run to its natural
+						throttleReasons[s.idx] = reason
-					// end so the daemon releases the slot cleanly before we
+						logFunc(fmt.Sprintf("power calibration: GPU %d detected %s throttle at %d W, waiting for run to finish", s.idx, reason, s.appliedLimitW))
-					// reduce power and retry.
-					if reason := benchmarkCalibrationThrottleReason(beforeThrottle, afterThrottle); reason != "" && throttleReason == "" {
-						throttleReason = reason
-						logFunc(fmt.Sprintf("power calibration: GPU %d detected %s throttle at %d W, waiting for current run to finish before reducing power limit", idx, reason, appliedLimitW))
 					}
-				case <-ctx.Done():
-					cancel()
-					attempt = <-doneCh
-					break attemptLoop
 				}
+			case <-ctx.Done():
+				cancelAttempt()
+				ar = <-doneCh
+				break attemptLoop
 			}
-			ticker.Stop()
+		}
-			cancel()
+		ticker.Stop()
-			_ = os.WriteFile(filepath.Join(runDir, logName), attempt.out, 0644)
+		cancelAttempt()
+		_ = os.WriteFile(filepath.Join(runDir, logName), ar.out, 0644)
 
-			perGPU := filterRowsByGPU(attempt.rows, idx)
+		// Resource busy: retry with exponential back-off (shared — one DCGM session).
+		if ar.err != nil && isDCGMResourceBusy(ar.err) {
+			if busyDelaySec > dcgmResourceBusyMaxDelaySec {
+				for _, s := range active {
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("DCGM resource busy after %d retries, giving up", busyRetries))
+					s.converged = true
+				}
+				logFunc(fmt.Sprintf("power calibration: DCGM resource persistently busy after %d retries, stopping", busyRetries))
+				break calibDone
+			}
+			busyRetries++
+			// Undo attempt counter: busy retries don't count as real attempts.
+			for _, s := range active {
+				s.calib.Attempts--
+			}
+			logFunc(fmt.Sprintf("power calibration: DCGM resource busy (attempt %d), retrying in %ds", sharedAttempt, busyDelaySec))
+			select {
+			case <-ctx.Done():
+				break calibDone
+			case <-time.After(time.Duration(busyDelaySec) * time.Second):
+			}
+			next := busyDelaySec * 2
+			if next > dcgmResourceBusyMaxDelaySec {
+				next = dcgmResourceBusyMaxDelaySec + 1
+			}
+			busyDelaySec = next
+			sharedAttempt-- // retry same logical attempt number
+			continue
+		}
+		busyRetries = 0
+		busyDelaySec = 1
+
+		// Per-GPU analysis and binary search update.
+		for _, s := range active {
+			perGPU := filterRowsByGPU(ar.rows, s.idx)
 			summary := summarizeBenchmarkTelemetry(perGPU)
-			if throttleReason == "" && attempt.err == nil && summary.P95PowerW > 0 {
+			throttle := throttleReasons[s.idx]
-				// Stable at appliedLimitW: record it and binary-search upward.
+
-				calib.Summary = summary
+			// Cooling warning: thermal throttle with fans not at maximum.
-				calib.Completed = true
+			if strings.Contains(throttle, "thermal") && s.calib.CoolingWarning == "" {
-				calib.AppliedPowerLimitW = float64(appliedLimitW)
+				clocks := make([]float64, 0, len(perGPU))
-				logFunc(fmt.Sprintf("power calibration: GPU %d stable at %d W, p95=%.0f W p95_temp=%.1f C (%d samples)", idx, appliedLimitW, summary.P95PowerW, summary.P95TempC, summary.Samples))
+				var fanDutyValues []float64
-				lo = appliedLimitW
+				fanDutyAvail := false
-				// If there is still headroom to search, try a higher level.
+				for _, r := range perGPU {
-				if canDerate && hi-lo > calibSearchTolerance {
+					if r.ClockMHz > 0 {
-					nextLimitW := roundTo5W((lo + hi) / 2)
+						clocks = append(clocks, r.ClockMHz)
-					if nextLimitW > lo && nextLimitW < hi {
+					}
-						if err := setBenchmarkPowerLimit(ctx, verboseLog, idx, nextLimitW); err == nil {
+					if r.FanDutyCycleAvailable {
-							appliedLimitW = nextLimitW
+						fanDutyAvail = true
-							calib.AppliedPowerLimitW = float64(appliedLimitW)
+						fanDutyValues = append(fanDutyValues, r.FanDutyCyclePct)
-							calib.Notes = append(calib.Notes, fmt.Sprintf("binary search: stable at %d W, trying %d W (lo=%d hi=%d)", lo, nextLimitW, lo, hi))
+					}
-							logFunc(fmt.Sprintf("power calibration: GPU %d binary search up: stable at %d W, trying %d W", idx, lo, nextLimitW))
+				}
-							continue calibLoop
+				dropPct := benchmarkClockDrift(clocks)
+				p95FanDuty := benchmarkPercentile(fanDutyValues, 95)
+				if dropPct >= 20 && fanDutyAvail && p95FanDuty < 98 {
+					s.calib.CoolingWarning = fmt.Sprintf(
+						"thermal throttle (%s) caused a %.0f%% clock drop while fans were at %.0f%% duty cycle — server cooling may not be configured for full GPU load",
+						throttle, dropPct, p95FanDuty,
+					)
+					logFunc(fmt.Sprintf("power calibration: GPU %d cooling warning: %s", s.idx, s.calib.CoolingWarning))
+				}
+			}
+
+			if throttle == "" && ar.err == nil && summary.P95PowerW > 0 {
+				// Stable at current limit — update lo and binary-search upward.
+				s.calib.Summary = summary
+				s.calib.Completed = true
+				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
+				if canDerate && s.hi-s.lo > calibSearchTolerance {
+					next := roundTo5W((s.lo + s.hi) / 2)
+					if next > s.lo && next < s.hi {
+						if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, next); err == nil {
+							s.appliedLimitW = next
+							s.calib.AppliedPowerLimitW = float64(next)
+							s.calib.Completed = false // keep searching
+							s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: stable at %d W, trying %d W (lo=%d hi=%d)", s.lo, next, s.lo, s.hi))
+							logFunc(fmt.Sprintf("power calibration: GPU %d binary search up: stable at %d W, trying %d W", s.idx, s.lo, next))
+							continue // next GPU in active list
 						}
 					}
 				}
-				break
+				s.converged = true
+				continue
 			}
 
-			// If DCGM reports the resource is in use, nv-hostengine has not yet
+			// Failed or throttled — log and binary-search downward.
-			// released the diagnostic slot from the previous attempt. Do not
-			// derate: wait with exponential back-off and retry at the same
-			// power limit. Once the back-off delay would exceed
-			// dcgmResourceBusyMaxDelaySec, fail — the slot is persistently
-			// held by something else.
-			if attempt.err != nil && isDCGMResourceBusy(attempt.err) {
-				if busyDelaySec > dcgmResourceBusyMaxDelaySec {
-					calib.Notes = append(calib.Notes, fmt.Sprintf("DCGM resource busy after %d retries, giving up", busyRetries))
-					logFunc(fmt.Sprintf("power calibration: GPU %d DCGM resource persistently busy after %d retries, stopping", idx, busyRetries))
-					break
-				}
-				busyRetries++
-				logFunc(fmt.Sprintf("power calibration: GPU %d DCGM resource busy (attempt %d), retrying in %ds", idx, calib.Attempts, busyDelaySec))
-				select {
-				case <-ctx.Done():
-					break calibLoop
-				case <-time.After(time.Duration(busyDelaySec) * time.Second):
-				}
-				next := busyDelaySec * 2
-				if next > dcgmResourceBusyMaxDelaySec {
-					next = dcgmResourceBusyMaxDelaySec + 1 // sentinel: next busy → fail
-				}
-				busyDelaySec = next
-				continue calibLoop
-			}
-			busyRetries = 0    // reset on any non-busy outcome
-			busyDelaySec = 1   // reset back-off
-
 			switch {
-			case throttleReason != "":
+			case throttle != "":
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power was canceled on attempt %d after %s throttling at %d W", calib.Attempts, throttleReason, appliedLimitW))
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d: %s throttle at %d W", s.calib.Attempts, throttle, s.appliedLimitW))
-				logFunc(fmt.Sprintf("power calibration: GPU %d throttled (%s) at %d W, reducing power limit", idx, throttleReason, appliedLimitW))
+				logFunc(fmt.Sprintf("power calibration: GPU %d throttled (%s) at %d W, reducing power limit", s.idx, throttle, s.appliedLimitW))
-				// Check whether the thermal throttle coincided with fans below
+			case ar.err != nil:
-				// maximum: that combination suggests cooling misconfiguration
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d failed at %d W: %v", s.calib.Attempts, s.appliedLimitW, ar.err))
-				// rather than a fundamental power-delivery limit.
+				logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power failed at %d W: %v", s.idx, s.appliedLimitW, ar.err))
-				if strings.Contains(throttleReason, "thermal") && calib.CoolingWarning == "" {
-					clocks := make([]float64, 0, len(perGPU))
-					var fanDutyValues []float64
-					fanDutyAvail := false
-					for _, r := range perGPU {
-						if r.ClockMHz > 0 {
-							clocks = append(clocks, r.ClockMHz)
-						}
-						if r.FanDutyCycleAvailable {
-							fanDutyAvail = true
-							fanDutyValues = append(fanDutyValues, r.FanDutyCyclePct)
-						}
-					}
-					dropPct := benchmarkClockDrift(clocks)
-					p95FanDuty := benchmarkPercentile(fanDutyValues, 95)
-					if dropPct >= 20 && fanDutyAvail && p95FanDuty < 98 {
-						calib.CoolingWarning = fmt.Sprintf(
-							"thermal throttle (%s) caused a %.0f%% clock drop while fans were at %.0f%% duty cycle — server cooling may not be configured for full GPU load",
-							throttleReason, dropPct, p95FanDuty,
-						)
-						logFunc(fmt.Sprintf("power calibration: GPU %d cooling warning: %s", idx, calib.CoolingWarning))
-					}
-				}
-			case attempt.err != nil:
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power attempt %d failed at %d W: %v", calib.Attempts, appliedLimitW, attempt.err))
-				logFunc(fmt.Sprintf("power calibration: GPU %d targeted_power failed at %d W: %v", idx, appliedLimitW, attempt.err))
 			default:
-				calib.Notes = append(calib.Notes, fmt.Sprintf("targeted_power attempt %d at %d W produced no valid power telemetry", calib.Attempts, appliedLimitW))
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("targeted_power attempt %d at %d W: no valid power telemetry", s.calib.Attempts, s.appliedLimitW))
-				logFunc(fmt.Sprintf("power calibration: GPU %d attempt %d at %d W produced no valid telemetry", idx, calib.Attempts, appliedLimitW))
+				logFunc(fmt.Sprintf("power calibration: GPU %d attempt %d at %d W: no valid telemetry", s.idx, s.calib.Attempts, s.appliedLimitW))
 			}
 
-			if !canDerate || appliedLimitW <= 0 {
+			if !canDerate || s.appliedLimitW <= 0 {
-				break
+				s.converged = true
+				continue
 			}
-			// Binary-search for the highest stable power limit.
+			s.hi = s.appliedLimitW
-			// This attempt failed or throttled, so update the upper bound.
-			hi = appliedLimitW
 
-			if hi-lo <= calibSearchTolerance {
+			if s.hi-s.lo <= calibSearchTolerance {
-				// Search range exhausted: lo is the highest verified-stable level.
+				if s.lo > s.minLimitW {
-				if lo > minLimitW {
+					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))
-					calib.Notes = append(calib.Notes, fmt.Sprintf("binary search converged: using %d W (lo=%d hi=%d)", lo, lo, hi))
+					if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, s.lo); err == nil {
-					if err := setBenchmarkPowerLimit(ctx, verboseLog, idx, lo); err == nil {
+						s.appliedLimitW = s.lo
-						appliedLimitW = lo
+						s.calib.AppliedPowerLimitW = float64(s.lo)
-						calib.AppliedPowerLimitW = float64(lo)
+						s.calib.Derated = s.lo < s.originalLimitW
-						calib.Derated = lo < originalLimitW
 					}
 				} else {
-					calib.Notes = append(calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
+					s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
 				}
-				break
+				s.converged = true
+				continue
 			}
 
-			// Compute the next candidate.
+			next := roundTo5W((s.lo + s.hi) / 2)
-			// For thermal throttle: use the pre-throttle power draw from telemetry
+			if next <= s.lo {
-			// as a smarter initial estimate instead of the binary midpoint — it
+				next = s.lo + calibSearchTolerance
-			// lands much closer to the true limit on the first attempt.
-			nextLimitW := (lo + hi) / 2
-			if strings.Contains(throttleReason, "thermal") {
-				if onsetW := calibPreThrottlePowerW(perGPU); onsetW > 0 {
-					candidate := roundTo5W(int(math.Round(onsetW)) - calibPreThrottleMarginW)
-					if candidate > lo && candidate < hi {
-						nextLimitW = candidate
-					}
-				}
 			}
-			nextLimitW = roundTo5W(nextLimitW)
+			if next >= s.hi {
-			// Ensure the candidate is strictly inside the search range.
+				next = (s.lo + s.hi) / 2
-			if nextLimitW <= lo {
-				nextLimitW = lo + calibSearchTolerance
 			}
-			if nextLimitW >= hi {
+			if next < s.minLimitW {
-				nextLimitW = (lo + hi) / 2
+				s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
+				s.converged = true
+				continue
 			}
-			if nextLimitW < minLimitW {
+			if err := setBenchmarkPowerLimit(ctx, verboseLog, s.idx, next); err != nil {
-				calib.Notes = append(calib.Notes, fmt.Sprintf("could not find a stable targeted_power limit within %d W of the default", maxDerateW))
+				s.calib.Notes = append(s.calib.Notes, "failed to set power limit: "+err.Error())
-				break
+				logFunc(fmt.Sprintf("power calibration: GPU %d failed to set power limit %d W: %v", s.idx, next, err))
+				s.converged = true
+				continue
 			}
-			if err := setBenchmarkPowerLimit(ctx, verboseLog, idx, nextLimitW); err != nil {
+			s.appliedLimitW = next
-				calib.Notes = append(calib.Notes, "failed to set power limit: "+err.Error())
+			s.calib.AppliedPowerLimitW = float64(next)
-				logFunc(fmt.Sprintf("power calibration: GPU %d failed to set power limit %d W: %v", idx, nextLimitW, err))
+			s.calib.Derated = next < s.originalLimitW
-				break
+			s.info.PowerLimitW = float64(next)
-			}
+			infoByIndex[s.idx] = s.info
-			appliedLimitW = nextLimitW
+			s.calib.Notes = append(s.calib.Notes, fmt.Sprintf("binary search: trying %d W (lo=%d hi=%d)", next, s.lo, s.hi))
-			calib.AppliedPowerLimitW = float64(appliedLimitW)
+			logFunc(fmt.Sprintf("power calibration: GPU %d binary search: trying %d W (lo=%d hi=%d)", s.idx, next, s.lo, s.hi))
-			calib.Derated = appliedLimitW < originalLimitW
-			info.PowerLimitW = float64(appliedLimitW)
-			infoByIndex[idx] = info
-			calib.Notes = append(calib.Notes, fmt.Sprintf("binary search: trying %d W (lo=%d hi=%d)", nextLimitW, lo, hi))
-			logFunc(fmt.Sprintf("power calibration: GPU %d binary search: trying %d W (lo=%d hi=%d)", idx, nextLimitW, lo, hi))
 		}
+	}
 
-		if calib.Completed || calib.Attempts > 0 || len(calib.Notes) > 0 {
+	for _, s := range states {
-			results[idx] = calib
+		if s.calib.Completed || s.calib.Attempts > 0 || len(s.calib.Notes) > 0 {
+			results[s.idx] = s.calib
 		}
 	}
 	return results, restore
@@ -2770,28 +2815,6 @@ func isDCGMResourceBusy(err error) bool {
 	return errors.As(err, &exitErr) && exitErr.ExitCode() == 222
 }
 
-// calibPreThrottlePowerW estimates the GPU power draw just before thermal
-// throttle onset by averaging the first quarter of telemetry rows. The early
-// samples capture the GPU at peak before clock/power reduction kicks in.
-func calibPreThrottlePowerW(rows []GPUMetricRow) float64 {
-	if len(rows) < 4 {
-		return 0
-	}
-	n := len(rows) / 4
-	var sum float64
-	var cnt int
-	for _, r := range rows[:n] {
-		if r.PowerW > 0 {
-			sum += r.PowerW
-			cnt++
-		}
-	}
-	if cnt == 0 {
-		return 0
-	}
-	return sum / float64(cnt)
-}
-
 // roundTo5W rounds w to the nearest 5 W boundary.
 func roundTo5W(w int) int {
 	return ((w + 2) / 5) * 5