inference/research
verifiedhigh · 0.88updated 2026-10-03

8x H100 node under real AI training draws only 76% of its 10.2 kW TDP rating; calibrated model reaches 11.4% MAPE versus 27-37% for TDP-based estimates

76.0 percent of rated 10.2 kW node TDP actually drawn under compute-intensive training load

Fields

id
benchmarks/h100-8gpu-node-loaded-76-percent-of-tdp-measured
Metric
quality
Value
76
Unit
percent of rated 10.2 kW node TDP actually drawn under compute-intensive training load
Methodology
MEASURED at Brookhaven National Laboratory on 8-GPU H100 systems during production AI training, not a synthetic loop: power was measured directly and combined with open-source benchmarking data to fit statistical models relating computational intensity to node-level power consumption. This is a TRAINING result, not an inference result - it is recorded here as evidence about the TDP-vs-actual gap, and it must not be read as an inference idle-draw figure. Metric 'quality' is a placeholder; the schema enum has no energy or power member.
Measured by
third_party
Reproducible
yes
notes
THE MOST TRANSFERABLE POWER NUMBER IN THIS TASK, and it is about method rather than a specific part: NAMEPLATE WATTS ARE NOT A POWER BUDGET. Even compute-intensive training on a 10.2 kW-rated 8xH100 node drew only 76% of the rating, and modelling from TDP gave 27-37% error where a calibrated model gave 11.4%. Combined with the repo's existing H200 finding (whole-node plateau 5806.9 W against a GPU-only nameplate sum of 5600 W - legitimately above it, because the node also carries two Xeon 8480C hosts, 2 TB DRAM, ten 400 GbE NICs and PSU losses), the rule is: TDP UNDERSTATES a loaded multi-GPU node, because the node's non-accelerator overhead is not in the GPU nameplate, while TDP OVERSTATES what a real workload pulls. Any cost-per-token model that multiplies tokens by a datasheet wattage is wrong in both directions at once. The paper also finds distinct power signatures for transformer vs convolutional architectures, with transformers showing fluctuations relevant to grid stability - which means a short measurement window can misattribute a ramp as a steady state (see the repo's pwr-short-run-measures-ramp-not-steady-state gotcha). The authors separately note that the 2025 LBNL update treats the underlying AI-server power uncertainty as resolved by measured and simulated data.

Notes

THE MOST TRANSFERABLE POWER NUMBER IN THIS TASK, and it is about method rather than a specific part: NAMEPLATE WATTS ARE NOT A POWER BUDGET. Even compute-intensive training on a 10.2 kW-rated 8xH100 node drew only 76% of the rating, and modelling from TDP gave 27-37% error where a calibrated model gave 11.4%. Combined with the repo's existing H200 finding (whole-node plateau 5806.9 W against a GPU-only nameplate sum of 5600 W - legitimately above it, because the node also carries two Xeon 8480C hosts, 2 TB DRAM, ten 400 GbE NICs and PSU losses), the rule is: TDP UNDERSTATES a loaded multi-GPU node, because the node's non-accelerator overhead is not in the GPU nameplate, while TDP OVERSTATES what a real workload pulls. Any cost-per-token model that multiplies tokens by a datasheet wattage is wrong in both directions at once. The paper also finds distinct power signatures for transformer vs convolutional architectures, with transformers showing fluctuations relevant to grid stability - which means a short measurement window can misattribute a ramp as a steady state (see the repo's pwr-short-run-measures-ramp-not-steady-state gotcha). The authors separately note that the 2025 LBNL update treats the underlying AI-server power uncertainty as resolved by measured and simulated data.

Cross-references

Accelerators (1)

Engine (1)