MLPerf v6.1 Dell 2x Xeon 6980P: whisper INT8 Offline aggregate samples/s

MLPerf Inference v6.1 closed division, CPU-only (summary.csv accelerator_model_name is N/A). Values copied from summary.csv, not from a per-system mlperf_log_summary.txt. AGGREGATE throughput for the whole node, NOT per-request decode and NOT per-socket. Prompt length, output length, and batch size are not in the summary row. weight_data_types INT8. Framework column is PyTorch. Scenario: Offline. System: Dell PowerEdge R770AP, 2x Xeon 6980P, 24x96GB DDR5-6400. Whisper, not a token-generation model. Units in summary.csv are Samples/s, stored under tps_aggregate because that is the aggregate-throughput member of the metric enum; the unit field says samples/s so it is not read as tokens. DIMM population is stated: 24x96GB at 6400 MT/s, which is 12 channels per socket at 1 DPC. Framework PyTorch, INT8. Do not compare this number to the llama3.1-8b tok/s rows. measured_by vendor: these are submitter results inside the MLCommons closed division (Intel, Dell, Supermicro, Quanta, Red Hat, Cisco), not an independent re-run.

Whisper, not a token-generation model. Units in summary.csv are Samples/s, stored under tps_aggregate because that is the aggregate-throughput member of the metric enum; the unit field says samples/s so it is not read as tokens. DIMM population is stated: 24x96GB at 6400 MT/s, which is 12 channels per socket at 1 DPC. Framework PyTorch, INT8. Do not compare this number to the llama3.1-8b tok/s rows.