News 3 min read

NVIDIA Vera Rubin debuts in MLPerf with up to 2x GB300 training speed


NVIDIA’s Vera Rubin platform made its first appearance in the MLPerf benchmarks on November 16, 2026. In MLPerf Training v6.1, a Vera Rubin NVL72 system had the fastest time to train on six of the seven benchmarks and up to 2x the performance per GPU of NVIDIA’s own GB300 NVL72, and in the first round of the new MLPerf Endpoints inference benchmark NVIDIA claims more than 10x the throughput per watt on agentic workloads. They are the first MLPerf numbers for the platform that follows Blackwell.

What’s new

From NVIDIA’s MLPerf announcement and its training technical blog:

  • Every benchmark covered. NVIDIA says it was the only platform with results on all seven MLPerf Training v6.1 benchmarks, which span large language models, image generation, recommender systems and reinforcement learning post-training.
  • DeepSeek-V3 671B in 4.14 minutes. On the hardest workload, a 671-billion-parameter Mixture-of-Experts model, Vera Rubin NVL72 trained to the target accuracy in 4.14 minutes on 1,024 GPUs across 16 racks. GB300 NVL72 took 8.03 minutes with the same 1,024 GPUs, which is where the “about 2x per GPU” figure comes from. On the other benchmarks NVIDIA puts the gain at 1.4x to 1.6x.
  • Scaling. Going from 256 to 1,024 Vera Rubin GPUs on DeepSeek-V3 kept above 97% scaling efficiency, cutting time to train from 16 minutes to about four.
  • The fastest absolute result was not Vera Rubin. A 10,240-GPU GB300 cluster ran DeepSeek-V3 in 1.38 minutes, the round’s best time. Microsoft Azure also submitted at 10,240 GPUs, on GB200 systems.
  • A new post-training benchmark. v6.1 adds MLPerf’s first reinforcement learning post-training test, Qwen 3.5 397B trained with GRPO. Vera Rubin led it with 10.44 minutes on 1,024 GPUs, submitting with both TensorRT-LLM and vLLM as the inference backend for the rollouts.

MLPerf Endpoints: agentic inference per watt

MLPerf Endpoints v1.0 is a new benchmark that sends agentic coding and enterprise workflows to systems through API endpoints and checks that the generated tool calls still match reference answers. NVIDIA’s Endpoints technical blog says that on Kimi K3, Vera Rubin NVL72 delivered more than 10x the throughput per kilowatt of GB300 NVL72 at a minimum of 50 tokens per second per user. Across Kimi K3, DeepSeek-V4.1-Flash and Qwen3.6-35B-A3B it reports 5x to 10x the throughput per kilowatt.

One caveat matters here: the Vera Rubin Endpoints results are provisional. NVIDIA’s footnote says they are pending MLCommons verification and may change after peer review, while the GB300 results it compares them with are verified.

What it means for developers

Most developers will meet Vera Rubin as a cloud instance, not a rack, and MLPerf is the closest thing to a neutral yardstick for what that instance buys. The useful reading of this round:

  • The gain depends on the model. The 2x figure is for a huge Mixture-of-Experts model whose training is dominated by GPU-to-GPU communication. For the other six benchmarks, including Llama 2 70B LoRA fine-tuning, NVIDIA measured 1.4x to 1.6x, not 2x.
  • Your framework is already on it. NVIDIA submitted the GRPO run with vLLM as well as TensorRT-LLM, so a post-training pipeline built on open-source tooling is not locked out of the new hardware.
  • Software keeps moving. NVIDIA also published unverified reruns, made after the submission deadline with a newer software build, that are faster again on Llama 3.1 8B, gpt-oss 20B and DLRMv4. Treat those as a hint of what driver and library updates bring, not as MLPerf results.
  • For inference, compare per watt and per user. Endpoints reports throughput per kilowatt at a fixed responsiveness, which maps to the trade-off you make when sizing an agent service: more users per GPU, or faster replies for each one.

The agent infrastructure NVIDIA is building around these chips includes the sandbox runtime it gave to the CNCF last week.

How it compares

MLPerf rounds are open to every vendor, but NVIDIA’s posts compare Vera Rubin only with NVIDIA’s previous generation, GB300 NVL72, and say nothing about other vendors’ submissions. The full tables, with every submitter and system, are on MLCommons for Training v6.1 and Endpoints. NVIDIA says 20 partner organizations submitted on its platform this round, including ASUS, Cisco, CoreWeave, Dell Technologies, Google Cloud, HPE, Lambda, Nebius, Oracle and Supermicro.

Sources