GPU Power Experiment Results

AI Datacenter Power Profile Analysis — Transfer Package

12 Experiment Groups

31 Total Runs

Hardware: NVIDIA RTX A6000 / PRO 6000

Sampling: 10 ms (NVML)

Generated: 2026-02-19 13:10

Quick Reference — Column Definitions

Idle Avg (W): Average power during pre-experiment idle phase (GPU loaded but no computation).
Train Avg / Peak (W): Average and peak power during model training compute (forward + backward pass).
Val Avg (W): Average power during validation (forward pass only, no gradients).
Infer Idle (W): Average power between inference requests ("hot idle" — model stays loaded).
Energy (J): Total energy consumed across all phases of the run.

#00 Baseline Reference

Standard conditions with no power throttling. Control experiment for comparison.

Variable: None (control)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet50	64	fixed	52.7	134.3	178.0	89.9	78.1	29670	📊 Interactive 🖼️ Static

#00 Smoke Test

Quick end-to-end validation of the measurement pipeline.

Variable: Pipeline validation

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	8	fixed	52.2	125.7	137.2	123.6	23.7	9149	📊 Interactive 🖼️ Static

#01 Batch Size Sweep

How does batch size affect power draw? Larger batches → higher GPU utilization → different power profiles.

Variable: Batch size (16 / 64 / 128)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet50	128	fixed	51.8	138.1	211.8	98.1	78.2	29205	📊 Interactive 🖼️ Static
resnet50	16	fixed	53.3	139.9	152.2	132.5	77.9	32440	📊 Interactive 🖼️ Static
resnet50	64	fixed	52.5	133.4	173.2	83.7	78.3	29732	📊 Interactive 🖼️ Static

#02 Operator Control (Variable Inference)

Non-uniform inference request patterns, simulating real operator traffic.

Variable: Variable inference interval

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	variable	53.7	128.5	152.2	133.8	78.2	34124	📊 Interactive 🖼️ Static

#03 Power Cap Sweep

Effect of nvidia-smi power limits on performance and power. Tests whether power capping is effective.

Variable: Power cap (345W / 460W / 575W)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	fixed	9.9	111.6	148.0	148.6	77.9	27225	📊 Interactive 🖼️ Static
resnet18	64	fixed	9.4	118.1	147.3	145.8	78.1	27206	📊 Interactive 🖼️ Static
resnet18	64	fixed	52.1	102.0	150.0	79.7	77.9	28881	📊 Interactive 🖼️ Static

#04 Stepwise Power Ramp

Gradually increasing/decreasing power limits to observe GPU adaptation behavior.

Variable: Ramp enable

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	fixed	54.1	105.8	148.3	79.3	77.9	29572	📊 Interactive 🖼️ Static

#05 Inference Pattern Comparison

Side-by-side comparison of three inference scheduling patterns and their grid impact.

Variable: Pattern (fixed / variable / burst)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	burst	51.9	102.0	151.7	79.5	78.4	28671	📊 Interactive 🖼️ Static
resnet18	64	fixed	53.4	100.1	149.9	80.6	78.3	28981	📊 Interactive 🖼️ Static
resnet18	64	variable	52.3	103.6	150.7	79.5	78.0	28852	📊 Interactive 🖼️ Static

#06 Model Architecture Scaling

How different AI model architectures produce different power signatures. Models range from lightweight (~55W) to heavy (~310W).

Variable: Model architecture (10 models)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
convnext_tiny	64	fixed	53.6	325.6	514.4	264.8	79.1	42655	📊 Interactive 🖼️ Static
densenet121	64	fixed	53.7	378.6	436.7	260.1	79.2	43719	📊 Interactive 🖼️ Static
efficientnet_b0	64	fixed	53.2	310.7	401.6	166.6	78.9	38061	📊 Interactive 🖼️ Static
mobilenetv2	64	fixed	54.0	103.0	138.9	81.8	78.0	29468	📊 Interactive 🖼️ Static
resnet18	64	fixed	52.9	101.9	147.4	80.7	77.9	28871	📊 Interactive 🖼️ Static
resnet50	64	fixed	53.7	134.7	173.4	92.4	78.0	29658	📊 Interactive 🖼️ Static
resnext50_32x4d	64	fixed	53.4	381.1	514.3	287.4	79.5	45881	📊 Interactive 🖼️ Static
swin_t	64	fixed	51.8	481.1	512.4	452.1	80.2	45197	📊 Interactive 🖼️ Static
vgg16	64	fixed	52.1	218.4	287.3	95.9	78.0	31082	📊 Interactive 🖼️ Static
vit_b_16	64	fixed	51.7	562.2	585.8	509.5	80.2	48888	📊 Interactive 🖼️ Static

#07 Dataset Comparison

Whether the dataset (image resolution, number of classes) changes the power profile.

Variable: Dataset (CIFAR-10 / CIFAR-100 / ImageNet)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet50	64	fixed	52.4	131.6	170.0	85.2	78.2	29700	📊 Interactive 🖼️ Static
resnet50	64	fixed	53.7	135.0	172.6	94.8	78.3	29821	📊 Interactive 🖼️ Static
resnet50	64	fixed	52.0	479.7	490.5	163.7	79.0	338171	📊 Interactive 🖼️ Static

#08 Training Mode Comparison

Fixed SGD training vs Optuna-based automated hyperparameter search.

Variable: Training mode (fixed / automl)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	fixed	54.2	102.5	147.7	82.2	77.9	28945	📊 Interactive 🖼️ Static

#09 Diffusion Models

Stable Diffusion (SD1.5 / SDXL) power experiments.

Variable: Diffusion model

No completed runs in this experiment group.

#10 Checkpoint I/O Observation

Periodic model saves create visible power spikes as GPU compute pauses for disk I/O.

Variable: Checkpoint frequency

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	fixed	53.5	—	—	79.4	77.9	29034	📊 Interactive 🖼️ Static

#11 Clock Frequency Sweep

Effect of locking GPU SM clocks at different frequencies. Lower clocks → less power, slower execution.

Variable: Clock lock (1005 / 1500 / 2100 MHz)

Model	Batch	Pattern	Idle Avg (W)	Train Avg (W)	Train Peak (W)	Val Avg (W)	Infer Idle (W)	Energy (J)	Plots
resnet18	64	fixed	52.9	64.4	79.1	54.1	53.5	21503	📊 Interactive 🖼️ Static
resnet18	64	fixed	57.9	70.9	88.1	59.3	57.9	23260	📊 Interactive 🖼️ Static
resnet18	64	fixed	61.2	76.3	98.1	63.1	61.8	24797	📊 Interactive 🖼️ Static

Master Summary — All Runs

Group	Model	Batch	Pattern	Idle (W)	Train Avg (W)	Train Peak (W)	Val (W)	Infer Idle (W)	Energy (J)	Plot
Baseline Reference	resnet50	64	fixed	52.7	134.3	178.0	89.9	78.1	29670	📊
Smoke Test	resnet18	8	fixed	52.2	125.7	137.2	123.6	23.7	9149	📊
Batch Size Sweep	resnet50	128	fixed	51.8	138.1	211.8	98.1	78.2	29205	📊
Batch Size Sweep	resnet50	16	fixed	53.3	139.9	152.2	132.5	77.9	32440	📊
Batch Size Sweep	resnet50	64	fixed	52.5	133.4	173.2	83.7	78.3	29732	📊
Operator Control (Variable Inference)	resnet18	64	variable	53.7	128.5	152.2	133.8	78.2	34124	📊
Power Cap Sweep	resnet18	64	fixed	9.9	111.6	148.0	148.6	77.9	27225	📊
Power Cap Sweep	resnet18	64	fixed	9.4	118.1	147.3	145.8	78.1	27206	📊
Power Cap Sweep	resnet18	64	fixed	52.1	102.0	150.0	79.7	77.9	28881	📊
Stepwise Power Ramp	resnet18	64	fixed	54.1	105.8	148.3	79.3	77.9	29572	📊
Inference Pattern Comparison	resnet18	64	burst	51.9	102.0	151.7	79.5	78.4	28671	📊
Inference Pattern Comparison	resnet18	64	fixed	53.4	100.1	149.9	80.6	78.3	28981	📊
Inference Pattern Comparison	resnet18	64	variable	52.3	103.6	150.7	79.5	78.0	28852	📊
Model Architecture Scaling	convnext_tiny	64	fixed	53.6	325.6	514.4	264.8	79.1	42655	📊
Model Architecture Scaling	densenet121	64	fixed	53.7	378.6	436.7	260.1	79.2	43719	📊
Model Architecture Scaling	efficientnet_b0	64	fixed	53.2	310.7	401.6	166.6	78.9	38061	📊
Model Architecture Scaling	mobilenetv2	64	fixed	54.0	103.0	138.9	81.8	78.0	29468	📊
Model Architecture Scaling	resnet18	64	fixed	52.9	101.9	147.4	80.7	77.9	28871	📊
Model Architecture Scaling	resnet50	64	fixed	53.7	134.7	173.4	92.4	78.0	29658	📊
Model Architecture Scaling	resnext50_32x4d	64	fixed	53.4	381.1	514.3	287.4	79.5	45881	📊
Model Architecture Scaling	swin_t	64	fixed	51.8	481.1	512.4	452.1	80.2	45197	📊
Model Architecture Scaling	vgg16	64	fixed	52.1	218.4	287.3	95.9	78.0	31082	📊
Model Architecture Scaling	vit_b_16	64	fixed	51.7	562.2	585.8	509.5	80.2	48888	📊
Dataset Comparison	resnet50	64	fixed	52.4	131.6	170.0	85.2	78.2	29700	📊
Dataset Comparison	resnet50	64	fixed	53.7	135.0	172.6	94.8	78.3	29821	📊
Dataset Comparison	resnet50	64	fixed	52.0	479.7	490.5	163.7	79.0	338171	📊
Training Mode Comparison	resnet18	64	fixed	54.2	102.5	147.7	82.2	77.9	28945	📊
Checkpoint I/O Observation	resnet18	64	fixed	53.5	—	—	79.4	77.9	29034	📊
Clock Frequency Sweep	resnet18	64	fixed	52.9	64.4	79.1	54.1	53.5	21503	📊
Clock Frequency Sweep	resnet18	64	fixed	57.9	70.9	88.1	59.3	57.9	23260	📊
Clock Frequency Sweep	resnet18	64	fixed	61.2	76.3	98.1	63.1	61.8	24797	📊