NVIDIA's next-generation Vera Rubin platform has made its first appearance on MLPerf, and the debut was a clean sweep. Results published by MLCommons show Vera Rubin as the only platform with submissions on every MLPerf Training v6.1 benchmark this round, covering large language models, image generation, recommender systems and reinforcement learning post-training. On top of the sweep, NVIDIA says Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized GB300 NVL72 systems that led earlier rounds.
The per-GPU jump matters more than the headline sweep. Each MLPerf round is crowded with heavily tuned GB300 submissions from cloud providers and system vendors, so beating the previous flagship by a generational margin — rather than a tuning margin — is the claim that will draw scrutiny. NVIDIA cautions that the numbers are early and will keep climbing as its CUDA software stack is optimized, meaning customers who buy Rubin systems now should see performance rise through software updates.
Scaling behavior was another focus. On the DeepSeek-V3 training benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with roughly 97 percent efficiency: quadrupling the hardware cut training time from about 16 minutes to about four. NVIDIA also ran the round's largest submission on 10,240 GB300 GPUs, with Microsoft Azure matching that scale using 10,240 GB200 GPUs. Submissions came from 20 organizations, including CoreWeave, Crusoe, Lambda, Nebius, Oracle, Google Cloud, HPE, Dell and Supermicro.
The more consequential debut may be in inference. MLPerf Endpoints is a new benchmark category that scores systems on real agentic coding and enterprise workflows while checking that the agents actually complete their tasks correctly — a deliberate response to the way token-heavy, multi-step agent reasoning stresses data centers that are capped by power, not floor space. In a provisional result for Moonshot's Kimi K3, a 32-accelerator Vera Rubin NVL72 system delivered a system throughput of about 41,414 tokens per second at roughly 101 kW of operating power, versus about 2,210 tokens per second at 60 kW for a 48-accelerator GB300 NVL72 system at a comparable end-to-end interactivity of around 50 tokens per second per user — a figure NVIDIA presents as more than 10x throughput per watt.
That comparison needs two caveats. It is a vendor submission, and the Vera Rubin Endpoints result is explicitly provisional pending verification by MLCommons, while the GB300 figure it is measured against has been verified. Even discounted, the direction is clear: the competitive metric for AI infrastructure is shifting from peak training throughput to how much agentic work each watt can deliver at usable response speeds, and Rubin was architected around that framing.
The debut lands during a crowded week for NVIDIA. The company reported fiscal third-quarter revenue of $115.8 billion, up 103 percent year over year, just as interim CEO Colette Kress told investors the company would stay disciplined about backing neocloud customers. Rubin performance claims will be tested against demand: the platform is scheduled to ramp through 2027 as the successor to Blackwell, and every point of published throughput-per-watt advantage feeds directly into the pricing conversations NVIDIA is having with hyperscalers weighing AMD alternatives.
Vera Rubin also led independent benchmarks including SemiAnalysis AgentX and Artificial Analysis AgentPerf, per NVIDIA's announcement. Independent re-testing of the MLPerf Endpoints numbers — particularly the 10x-per-watt agentic claim — will be the number to watch once MLCommons completes its review.
Comments (0)
Log in to join the discussion
Log InNo comments yet