Post by CoreWeave

We just trained DeepSeek-V3 671B benchmark in 2 minutes. 671 billion parameters. 8,192 NVIDIA Blackwell Ultra GPUs. A 2 minute time-to-train, the fastest DeepSeek-V3 run ever recorded in MLPerf®. This wasn't a benchmark-only cluster built to win a leaderboard. This is the same CoreWeave Cloud infrastructure our customers train on every day. Read that again: This is the same CoreWeave Cloud infrastructure our customers train on every day. Here are the facts: ► We are the only team to scale a NVIDIA GB300 NVL72 platform past 2,048 GPUs on DeepSeek-V3 in this MLPerf round. ► Connected with NVIDIA Spectrum-X Ethernet, we doubled it to 4,096 GPUs, then doubled it again to 8,192 GPUs, and held strong scaling efficiency the whole way. ► From 8,192 NVIDIA Blackwell Ultra GPUs to 64 NVIDIA Blackwell GPUs, we delivered high performance consistently demonstrating that customers can obtain the same performance across all cluster sizes on CoreWeave Cloud.   Check out the final results. https://utm.io/uqCko