Source · Tier A · Peer-reviewed

Accurate and Convenient Energy Measurements for GPUs: A Detailed Study of NVIDIA GPU's Built-In Power Sensor

Z. Yang, K. Adamek, W. Armour. 2024. SC24: International Conference for High Performance Computing, Networking, Storage and Analysis.

Originalhttps://arxiv.org/abs/2312.02741
DOI10.1109/SC41406.2024.00028
arXiv2312.02741
VersionPublished in the SC24 proceedings (presented 2024-11-19; DOI 10.1109/SC41406.2024.00028). The arXiv preprint is titled "Part-time Power Measurements: nvidia-smi's Lack of Attention" (v1 2023-12-05, v3 2024-12-12). The arXiv v3 abstract and HTML were read on 2026-09-25; the SC24 programme abstract was read the same day and states the 25% sampling finding. The IEEE and ACM pages could not be read.
Accessed2026-09-25
NoteMicro-benchmark study of nvidia-smi power readings on over 70 NVIDIA GPUs across 12 architecture generations. States that nvidia-smi uses NVML (§2.4), that on A100 and H100 GPUs only 25% of runtime is sampled for power, and that the reading's error is ±5% (within ±5% in most cases, §4.2) rather than the ±5 W NVIDIA claims.

Cited by

Search

Full search page