The Perf Testing Behind CoreWeave's 2.02-Minute MLPerf Record

In 2026, CoreWeave published results in four MLPerf® rounds in six months: the fastest DeepSeek-V3 671B time-to-train in MLPerf® Training v6.0 (2.02 minutes at 8,192 GPUs, to MLPerf's defined quality target), leading per-GPU DeepSeek-R1 inference on NVIDIA GB200 NVL72, serving leadership under realistic concurrent load in MLPerf® Endpoints, and a fourth round landing just weeks before this talk. This session decomposes those headline numbers the way a skeptical buyer should: what each benchmark actually measures, what the multipliers are made of, and which caveats matter. Then it covers the part that never makes the press release: the continuous performance testing, pre-run cluster qualification, and topology-aware placement that informed all four results. You will leave with three questions to ask any provider about their benchmark claims.

Demo: https://coreweave.atlassian.net/wiki/spaces/CWBEAT/pages/1950090966/MLPerf+v6.0+Training+Under+the+Hood