NVIDIA announced on Nov. 16 that its Vera Rubin platform led the MLPerf Training v6.1 benchmark, with results released the same day. The results mark Vera Rubin's first appearance in MLPerf benchmarks, which measure AI training speed and inference throughput. MLPerf is run by MLCommons; vendors submit their own results, which MLCommons then reviews.
NVIDIA's announcement positions Vera Rubin as a performance improvement over its predecessor, the Blackwell Ultra architecture, for both training large models and running AI agents. NVIDIA platforms were the only ones with results on all seven MLPerf Training v6.1 benchmarks, and Vera Rubin completed the full suite, according to the NVIDIA Newsroom1. The company said Vera Rubin NVL72 delivered up to 2x more performance per GPU than the GB300 NVL72 system on the DeepSeek-V3 671B benchmark, training the model in 4.14 minutes using 1,024 GPUs, compared with 8.03 minutes for GB300 NVL72 at the same scale, according to an NVIDIA technical blog2. Vera Rubin NVL72 also led the new Qwen 3.5 397B GRPO reinforcement learning post-training benchmark with a 10.44-minute result at 1,024 GPUs.
For the DeepSeek-V3 benchmark, GB300 NVL72 produced the fastest absolute result of the round at 1.38 minutes on 10,240 GPUs. NVIDIA said 80% of closed-division submissions in the training round ran on NVIDIA platforms across 20 partner organizations.
In the MLPerf Endpoints v1.0 benchmark, released Nov. 9, 2026, NVIDIA reported provisional results showing Vera Rubin NVL72 delivering over 10x the throughput per kilowatt of GB300 NVL72 for the Kimi K3 model at a shared minimum end-to-end interactivity of 50 tokens per second per user, according to an NVIDIA technical blog3. The Vera Rubin result used 32 accelerators at 1,024 concurrent queries, while the GB300 result used 48 accelerators at 48 queries. Across three models, NVIDIA said Vera Rubin achieved 5x to 10x higher throughput per kilowatt. The Endpoints results are pending MLCommons verification.
NVIDIA attributed the performance gains to the Vera Rubin architecture, which combines 72 Rubin GPUs and 36 Vera CPUs in a single rack-scale system connected by a copper NVLink 6 spine delivering 3.6 TB/s per GPU. The company said the platform benefits from continuous software optimization. Unverified re-runs of several benchmarks with a newer software stack showed further reductions in training time.
The MLCommons result entries were not independently reviewed for this report. NVIDIA said Vera Rubin systems are ramping into full production, according to the company's product page4, and are generally available through cloud infrastructure partners worldwide. The announcement did not include independent tests or pricing.