MLPerf Inference v5.1 official results database, summary_results.json (1448 result rows, 24 power-measured)

The v5.1 round. 1448 rows, of which 24 carry has_power=true - and NONE of those 24 is an LLM: the entire power-measured population is stable-diffusion-xl, retinanet and dlrm. So no Llama result in v5.1 is power-instrumented either. v5.1 is however the richest round for DECLARED power configuration, because the Notes field is populated with per-accelerator TGP strings: 'B200 TGP 1000W', 'H200 TGP 700W', 'H200-NVL TGP 600W', 'L40S TGP 350W'. Also the source of a real units defect: eight llama3.1-8b Interactive rows are published with Performance_Units='Queries/s' while the underlying harness log for the same run prints 'Completed tokens per second' - see bench-mlperf-v5-1-supermicro-b200-llama3-8b-interactive-log.