MLPerf v6.1 CoreWeave GB200 NVL72 (72x GB200-186GB aarch64) llama2-70b-99.9 Server run_1 harness summary log

Fields

id

sources/bench-mlperf-v6-1-coreweave-gb200-nvl72-llama2-server-log

URL

https://raw.githubusercontent.com/mlcommons/inference_results_v6.1/main/closed/CoreWeave/results/GB200-NVL72_GB200-186GB_aarch64x72_TRT/llama2-70b-99.9/Server/performance/run_1/mlperf_log_summary.txt

Publisher

MLCommons

Kind

benchmark

Accessed

2026-10-04

notes

Result is : VALID. Completed samples per second 3353.29, Completed tokens per second 910621.36. Performance constraints satisfied: Yes. NOTE THE INVERSION: at rack scale Server (910621) EXCEEDS Offline (894002) on the same 72 GPUs. This is not a contradiction - at this scale the Server harness keeps every GPU fed across 18 nodes while the Offline run pays scheduling and teardown on a single enormous batch - and it means 'Offline is always the higher number' is false outside the single-node case.

Notes

Result is : VALID. Completed samples per second 3353.29, Completed tokens per second 910621.36. Performance constraints satisfied: Yes. NOTE THE INVERSION: at rack scale Server (910621) EXCEEDS Offline (894002) on the same 72 GPUs. This is not a contradiction - at this scale the Server harness keeps every GPU fed across 18 nodes while the Offline run pays scheduling and teardown on a single enormous batch - and it means 'Offline is always the higher number' is false outside the single-node case.

Cross-references

This record references no other records.