chips infrastructure
Thirty Percent Less Power Did Not Cost Every Training Job Equally
Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.
Summary
Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.
The study defines a Power Flexibility Index to measure how much LLM-training throughput changes when GPU power is reduced. Its evidence covers 131 H200 runs, 24 H200 validations and 34 matched H100 runs across dense and mixture-of-experts models, pretraining and fine-tuning, and deployments up to 32 GPUs. Under a 30% power reduction, allocating power by the learned index recovered about 1,500 tokens per second per job—63% of the gap between equal allocation and perfect foresight. The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts.
Why it matters
Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.
Limits and context
- The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts.
Key claims
Across 189 H100 and H200 runs, a job-aware allocator recovered 63% of the throughput gap to an oracle under a power cap.
Qualification: The result shows measurable job-level flexibility, not a general claim about data-center electricity or grid impacts.
Evidence: source-2026-09-11-008
Sources
- arXiv preprint 2609.11542arXiv · primary research
Corrections
No corrections have been recorded for this story.