# Gpu Power Cap Eval

> Evaluates the performance and power efficiency trade-offs of NVIDIA H100 and H200 GPUs under varying power caps, isolating compute-bound (DGEMM) and memory-bound (STriad) workloads to analyze architectural scaling and frequency throttling dynamics. Use when the user wants to benchmark on cuBLAS DGEMM, TheBandwidthBenchmark (STriad kernel), or asks about evaluating this task. Reports Throughput (TFlop/s or TB/s).

- Skill: `qhjqhj00/gpu-power-cap-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/gpu-power-cap-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/gpu-power-cap-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/gpu-power-cap-eval

---


# gpu-power-cap-eval

> Architectural Trade-offs in the Energy-Efficient Era: A Comparative Study of power-capping NVIDIA H100 and H200 — Aditya Ujeniya et al. (arXiv:2604.11391, 2026)

## What this evaluates

Evaluates the performance and power efficiency trade-offs of NVIDIA H100 and H200 GPUs under varying power caps, isolating compute-bound (DGEMM) and memory-bound (STriad) workloads to analyze architectural scaling and frequency throttling dynamics.

## Datasets

- **cuBLAS DGEMM** — total ?; splits: test (-1)
- **TheBandwidthBenchmark (STriad kernel)** — total ?; splits: test (-1); repo https://github.com/HPC-Dwarfs/TheBandwidthBenchmark

## Metrics

- `Throughput (TFlop/s or TB/s)` **(primary)** — range: other
  - Average sustained throughput calculated as total operations (DGEMM) or data moved (STriad) divided by execution time, averaged over 50 runs per power-cap setting.
- `Power Draw (W)` — range: other
  - Average total and memory power consumption recorded via nvidia-smi polling at 10-second intervals, time-averaged per run.

## Input / output format

**Input**: GPU variant (H100/H200), power cap setting (200–700 W in 100 W steps), benchmark kernel (DGEMM 32768×32768 or STriad 80 GB array), and random data initialization.

**Output**: Average sustained throughput (TFlop/s or TB/s), average total power draw (W), average memory power draw (W), average SM frequency (MHz), and average memory frequency (MHz) per power-cap setting.

## Scoring recipe

```python
results = {}
for cap in range(200, 701, 100):
    runs = []
    for _ in range(50):
        metrics = run_benchmark(cap) # captures throughput, W_total, W_mem, freq_sm, freq_mem
        runs.append(metrics)
    results[cap] = {k: np.mean([r[k] for r in runs]) for k in runs[0]}
```

## Common pitfalls

- Memory frequency remains strictly static across all power caps, contrary to typical frequency-scaling assumptions.
- Power draw measurements via nvidia-smi polling (10s interval) may miss transient power spikes, and initialization method (random vs constant) can cause up to 100 W variance.
- Throughput is reported as sustained average, not peak, which differs from standard hardware spec reporting.

## Evidence (verbatim from paper)

> We executed the two benchmarks independently across all GPUs on four nodes. For each execution, we capture the benchmark throughput (in TFlop/s or TB/s), along with average metrics for total power draw (W), memory power draw (W), SM frequency (MHz), and memory frequency (MHz). ... We evaluated both the NVIDIA H100 and H200 across a power-capping range from 200 W to their maximum Thermal Design Power (TDP) of 700 W in increments of 100 W.

## Citation

```bibtex
@misc{ujeniya2026architectural,
  title={Architectural Trade-offs in the Energy-Efficient Era: A Comparative Study of power-capping NVIDIA H100 and H200},
  author={Aditya Ujeniya et al.},
  year={2026},
  note={arXiv:2604.11391}
}
```

- arXiv: 2604.11391