MLPerf Inference v4.1 NVIDIA H200-SXM-141GBx8_TRT_MaxQ llama2-70b-99.9 Offline LoadGen summary

AUDITED MLPerf LoadGen result for the same run as the spl.txt power trace. Key fields verbatim: SUT name 'llama2-70b-Offline-SUT', Scenario Offline, Mode PerformanceOnly, 'Samples per second: 86.7369', 'Tokens per second: 25262.1', 'Result is : VALID', min_duration 2400000 ms, samples_per_query 253440, target_qps 96, ttft_latency 2000000000 ns, tpot_latency 200000000 ns, performance_sample_count 24576, and the harness reports 'No warnings encountered during test.' This is the throughput denominator for every joules/token figure derived from the paired spl.txt. NOTE ON WHAT THE TOKENS PER SECOND DENOMINATOR COUNTS: MLPerf's Offline 'Tokens per second' is aggregate over the measured window at the harness level; it is not a per-request decode rate and should not be compared against a single-stream decode tok/s record. 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.