diff --git a/audit/internal/platform/benchmark.go b/audit/internal/platform/benchmark.go
index 840e22c..a630707 100644
--- a/audit/internal/platform/benchmark.go
+++ b/audit/internal/platform/benchmark.go
@@ -4008,14 +4008,23 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 			idleW = result.ServerPower.IdleW
 		}
 
-		// Build header: Run | GPU 0 | GPU 1 | ... | Server wall W | Per GPU wall W | Platform eff.
+		// Build header: Run | GPU 0 | GPU 1 | ... | GPU total W | Server itself W | Server wall W | Per GPU wall W | Platform eff.
 		headers := []string{"Run"}
 		for _, idx := range allGPUIndices {
 			headers = append(headers, fmt.Sprintf("GPU %d W", idx))
 		}
-		headers = append(headers, "Server wall W", "Per GPU wall W", "Platform eff.")
+		headers = append(headers, "GPU total W", "Server itself W", "Server wall W", "Per GPU wall W", "Platform eff.")
 
 		var rampRows [][]string
+		if idleW > 0 {
+			idleRow := []string{"0 (idle)"}
+			for range allGPUIndices {
+				idleRow = append(idleRow, "—")
+			}
+			// No load: GPU total is negligible, all draw is the server's own baseline.
+			idleRow = append(idleRow, "—", fmt.Sprintf("%.0f", idleW), fmt.Sprintf("%.0f", idleW), "—", "—")
+			rampRows = append(rampRows, idleRow)
+		}
 		for _, step := range result.RampSteps {
 			row := []string{fmt.Sprintf("%d", step.StepIndex)}
 			for _, idx := range allGPUIndices {
@@ -4036,6 +4045,16 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				}
 				row = append(row, gpuPwr)
 			}
+			// GPU total W = sum of observed GPU power (nvidia-smi)
+			gpuTotal := "—"
+			if step.TotalObservedPowerW > 0 {
+				gpuTotal = fmt.Sprintf("%.0f", step.TotalObservedPowerW)
+			}
+			// Server itself W = server wall power minus GPU total (non-GPU baseline draw)
+			serverItself := "—"
+			if step.ServerLoadedW > 0 && step.TotalObservedPowerW > 0 {
+				serverItself = fmt.Sprintf("%.0f", step.ServerLoadedW-step.TotalObservedPowerW)
+			}
 			// Server wall W
 			serverWall := "—"
 			if step.ServerLoadedW > 0 {
@@ -4055,7 +4074,7 @@ func renderPowerBenchReport(result NvidiaPowerBenchResult) string {
 				}
 				platEff = fmt.Sprintf("%.2f", eff)
 			}
-			row = append(row, serverWall, perGPUWall, platEff)
+			row = append(row, gpuTotal, serverItself, serverWall, perGPUWall, platEff)
 			rampRows = append(rampRows, row)
 		}
 		b.WriteString(fmtMDTable(headers, rampRows))
@@ -4617,6 +4636,8 @@ func (s *System) RunNvidiaPowerBench(ctx context.Context, baseDir string, opts N
 			ramp.AvgFanRPM = singleRun.AvgFanRPM
 			ramp.AvgFanDutyCyclePct = singleRun.AvgFanDutyCyclePct
 		}
+		firstSummary := firstCalib.Summary
+		ramp.PerGPUTelemetry = map[int]*BenchmarkTelemetrySummary{firstIdx: &firstSummary}
 		if !firstCalib.Completed {
 			ramp.Status = "FAILED"
 			ramp.Notes = append(ramp.Notes, fmt.Sprintf("GPU %d did not complete single-card %s", firstIdx, benchmarkPowerEngineLabel(benchmarkPowerEngine())))