NVIDIA said its Vera Rubin NVL72 platform made its debut in MLPerf Training v6.1, posting the quickest training time on six of seven benchmarks. Compared with an optimized GB300 NVL72 baseline, it delivered up to 2x more performance per GPU. On the DeepSeek-V3 671B Mixture-of-Experts model, it trained in 4.14 minutes on 1,024 GPUs, while GB300 NVL72 took 8.03 minutes at that scale. The post is dated November 16, 2026.
For the DeepSeek-V3 model, NVIDIA reports that Vera Rubin NVL72 kept above 97% scaling efficiency when going from 256 to 1,024 GPUs. Its networking includes a copper NVLink 6 spine and ConnectX-9 SuperNICs. In the new Qwen 3.5 397B GRPO reinforcement learning post-training benchmark, Vera Rubin NVL72 recorded 10.44 minutes at 1,024 GPUs, and it submitted using both TensorRT-LLM and vLLM inference backends.
NVIDIA also shared unverified reruns on the same hardware with newer software, covering Llama 3.1 8B, gpt-oss 20B, and DLRMv4. These results are not verified by MLCommons Association and have not gone through MLPerf review. NVIDIA reports that 80 percent of closed-division entries ran on its platforms, with 20 partner organizations involved. NVIDIA says Vera Rubin NVL72 is broadly available via cloud infrastructure partners across the world.
Key facts
- DeepSeek-V3 671B finished in 4.14 minutes on Vera Rubin NVL72, spread over 16 racks and 1,024 GPUs.
- GB300 NVL72 took 8.03 minutes on the same 1,024-GPU setup.
- Qwen 3.5 397B GRPO result: 10.44 minutes at 1,024 GPUs, versus 13.78 minutes for GB300 NVL72.
- Scaling efficiency stayed above 97% going from 256 to 1,024 GPUs on DeepSeek-V3.
- The copper NVLink 6 spine provides 3.6 TB/s per GPU inside each rack.
Why it matters
For people who use or build AI, the results point to faster training times for large models and reinforcement learning post-training, which NVIDIA ties to agentic AI workflows. NVIDIA argues that cutting training time in half could roughly double the experiments a team runs in a quarter. Its software updates after shipping are described as ongoing gains, though the reruns are unverified and the results come from NVIDIA.




