Nvidia’s Vera Rubin NVL72 system topped all seven tests in the MLPerf Training v6.1 benchmarks on November 16, 2026, posting the fastest training times on six of them. It delivered up to twice the performance per GPU of the GB300 NVL72, including approximately 97% scaling efficiency between 256 and 1,024 GPUs on DeepSeek-V3 671B. Training DeepSeek-V3 with 1,024 Rubin GPUs was completed in 4.14 minutes. In MLPerf Endpoints v1.0, Vera Rubin also achieved more than 10 times the throughput per watt of the GB300 NVL72 on the Kimi K3 model, while CoreWeave reported a 2.7-times speed improvement over its previous GB200 NVL72 setup on DeepSeek-V3. The platform retains a rack configuration of 72 GPUs paired with 36 CPUs, with HBM4 memory and sixth-generation NVLink interconnects.
Source: Read the original article


