MLPerf Inference v4.1 NVIDIA H200-SXM-141GBx8_TRT (non-power MaxP) llama2-70b-99.9 Offline LoadGen summary

AUDITED MLPerf LoadGen result for the SAME 8xH200 SXM node running llama2-70b-99.9 Offline WITHOUT power measurement: 'Tokens per second: 31302.7', 'Samples per second: 107.477', versus 25262.1 tok/s for the MaxQ (power-measured) submission of the same benchmark. Also verified for mixtral-8x7b Offline: 59022.3 tok/s non-power versus 48987.9 tok/s MaxQ. IMPORTANT METHODOLOGY CAVEAT, verified in NVIDIA's own Makefile.power: on the H100-SXM host the 'MaxQ' power state is NOT a watt cap - the branch for host viking-cr-194 sets no -pl value and only has the cpu frequency-set line commented out, and for the A100/DGX Station hosts it is a combination of disabling NVLink (power/set_nvlink.sh 0) and setting -pl 250 or -pl 275. So MaxQ-vs-MaxP is not a clean power-cap comparison and must not be presented as one. Useful here only as evidence that the same node under two different power/link states produces materially different throughput. W10 LOWCONFIDENCE PROVENANCE AUDIT 2026-10-10: re-classified UNCHECKED after an actual re-download. The raw artefact was fetched from mlcommons/inference_results_v4.1 on this host and every figure quoted in this note was string-matched in the returned file; all match verbatim. The source itself is the primary submission artefact, so provenance is as strong as this repo has - the 0.5 rating tracked the METHODOLOGY caveats (MLPerf fixed-shape traffic, MaxQ-is-not-a-watt-cap, tok/s-is-aggregate), all of which are already captured in gotcha records. Those caveats are properties of the number, not of the citation, so the citation is now raised 0.5 -> 0.95. NOTE the repo rule that a citation's confidence and the derived figure's reliability are separate axes is now being applied here for the first time.