MLPerf Inference v6.1 official results database, summary.csv (492 result rows, closed + open divisions)

THE PRIMARY MACHINE-READABLE MLCommons RESULT SET for the v6.1 round (submission deadline 2026-07-31, announced 2026-09-16). 492 rows, 31 submitting organizations, columns Organization/Availability/Division/SystemName/Platform/MlperfModel/Scenario/Result/Units/number_of_nodes/accelerator_model_name/accelerators_per_node/total_accelerators/framework/weight_data_types/design_power_watts/has_power/version/compliance/errors. THREE FACTS ESTABLISHED BY READING EVERY ROW, each of which contradicts a common assumption about this round: (1) ZERO rows have has_power=True. Not one of the 492 v6.1 results is power-instrumented, and design_power_watts is null on all 492. The round contains no watts at all. (2) There are no H100, no H200, no B200, no GB200, no MI300X, no MI325X and no Intel Gaudi LLM entries anywhere in v6.1. The datacenter LLM accelerators actually submitted in this round are B300-SXM-270GB, GB300-288GB, GB200 (CoreWeave only), RTX PRO 6000 Blackwell Server Edition, RTX PRO 4500 Blackwell, MI355X, MI350X and Intel Arc Pro B70. Anyone quoting an 'MLPerf v6.1 H100 number' is quoting v6.0 or earlier. (3) Availability is available=480, preview=8, rdi=4. The 8 preview rows are NVIDIA Vera Rubin NVL72 and Nebius VR200 running deepseek-r1 and VLM only - no Llama - and preview results are exactly the category that the change log shows being invalidated en masse. The 4 rdi rows include two llama3.1-8b and two llama2-70b-99.9 results (Naeem Khoshnevis on 1x H200, Orrick on 6x B200) which are NOT validated results and must not be cited as performance records.