SummaryNVIDIA's Vera Rubin platform made its debut in the MLPerf Training v6.1 and Endpoints benchmarks. The company claims it achieved complete dominance in training categories and that agentic inference efficiency improved by more than 10x per watt compared to its previous GB300 generation. |
Why it mattersMLPerf results are one of the few standardized, third-party-verified benchmarks in AI infrastructure. Therefore, winning all categories on the first attempt signals that Vera Rubin is entering the market as the reference platform for large-scale training and agentic workloads. The emphasis on tokens per watt in multi-step agent tasks reflects a real shift in data center economics. The limiting factor is no longer just FLOPS, but power. |
Highlights- Vera Rubin NVL72 delivers up to a 2x increase in per-GPU performance compared to GB300 NVL72 in Training v6.1. It achieves approximately 97% efficiency when scaling from 256 to 1,024 GPUs on DeepSeek-V3.
- In the Kimi K3 agentic benchmark, Vera Rubin delivers over 10x better performance in watts per efficiency compared to GB300 at a common interaction threshold of 50 tokens/second/user.
- 20 partner organizations, including Azure, Google Cloud, CoreWeave, and Oracle, submitted training results. This highlights the breadth of the CUDA ecosystem.
|
Critical takeThe 10x agentic inference value in the headline is explicitly marked as 'provisional' and is awaiting MLCommons peer review. Furthermore, this comparison is a self-referential evaluation against NVIDIA's own previous generation, not against AMD, Intel, or custom silicon. Moreover, the 'complete dominance' in Training v6.1 stems from NVIDIA being the only vendor submitting results in all benchmark categories. This confuses scope integrity with dominance. |
This critique was written by the local AI model
Qwen3.8-27B. I chose the story; I do not review each critique before it is published.