benchmarkNVIDIA

NVIDIA Vera Rubin NVL72 posts fastest time to train on six of seven MLPerf Training v6.1 benchmarks

TL;DR

In its first MLPerf Training appearance, NVIDIA's Vera Rubin NVL72 delivered the fastest time to train on six of seven benchmarks, with up to 2x the per-GPU performance of an optimized GB300 NVL72 baseline, according to NVIDIA. On DeepSeek-V3 671B it trained in 4.14 minutes on 1,024 GPUs versus 8.03 minutes for GB300 NVL72.

3 min read
0

NVIDIA's Vera Rubin NVL72 delivered the fastest time to train on six of the seven benchmarks in MLPerf Training v6.1, achieving up to 2x more performance per GPU than an optimized GB300 NVL72 baseline, according to NVIDIA's analysis of the closed-division results. It is the first MLPerf Training appearance for the Vera Rubin architecture.

The headline result: DeepSeek-V3 671B

On the DeepSeek-V3 671B Mixture-of-Experts benchmark, Vera Rubin NVL72 trained the model in 4.14 minutes on 1,024 GPUs. GB300 NVL72 took 8.03 minutes at the same scale. NVIDIA says the largest gain lands on the largest, most communication-heavy model.

For this seventh benchmark, the absolute fastest time of the round came from a 10,240-GPU GB300 NVL72 cluster, per NVIDIA. NVIDIA says its platforms together hold the fastest time to train on every benchmark in v6.1.

Per-GPU gains across the suite

NVIDIA's comparison of Vera Rubin NVL72 against GB300 NVL72, per GPU on the same model and GPU count (closed division, results retrieved from mlcommons.org on Nov 16, 2026):

Benchmark Model type GPU count Vera Rubin vs. GB300
DeepSeek-V3 671B Mixture-of-Experts 1,024 2x
Qwen3.5 397B GRPO (RL) Mixture-of-Experts 64 1.4x
gpt-oss 20B Mixture-of-Experts 64 1.6x
Llama 2 70B LoRA Dense 72 1.5x
Llama 3.1 8B Dense 72 1.6x
FLUX.1 Diffusion 72 1.5x
DLRMv4 Recommender 128 1.4x

NVIDIA describes the gains outside DeepSeek-V3 as consistently 1.4x to 1.6x. Qwen3.5 397B GRPO is a new reinforcement learning post-training benchmark in this round. NVIDIA says Vera Rubin submitted it with both TensorRT-LLM and vLLM inference backends, which it cites as evidence of software maturity.

NVIDIA also says its platforms were the only ones to submit results on all seven benchmarks, and that Vera Rubin completed the entire suite in its first run.

Scaling and interconnect

Vera Rubin NVL72 maintained above 97% overall scaling efficiency when expanding from 256 to 1,024 GPUs on DeepSeek-V3, according to NVIDIA. The company attributes this to:

  • A copper NVLink 6 spine delivering 3.6 TB/s per GPU.
  • ConnectX-9 SuperNICs.
  • Quantum-X800 InfiniBand or Spectrum-X Ethernet networking.

Unverified reruns

NVIDIA also reports reruns with a newer software stack that further reduced Vera Rubin NVL72 time to train on Llama 3.1 8B, gpt-oss 20B and DLRMv4. These reruns are unverified, meaning they are not part of the official MLPerf results. NVIDIA presents them as evidence of continued optimization after hardware ships.

Ecosystem and availability

NVIDIA says eighty percent of closed-division submissions ran on NVIDIA platforms, across 20 partner organizations. Vera Rubin NVL72 systems are generally available through cloud infrastructure partners worldwide, according to NVIDIA. Pricing is not disclosed in the post.

What this means

The per-GPU comparison at matched GPU counts isolates the generational change from NVIDIA's own prior platform, and NVIDIA's reported figures put that change between 1.4x and 2x depending on the workload. The 2x result is on DeepSeek-V3 671B, the largest model in the suite, and NVIDIA ties the widest gap to communication-heavy training. The 97% scaling-efficiency figure covers 256 to 1,024 GPUs only, so it does not speak to larger Vera Rubin configurations. For that model, the fastest absolute time came from a 10,240-GPU GB300 NVL72 cluster rather than Vera Rubin. Teams planning training or RL post-training runs should read the published MLCommons entries, listed in NVIDIA's post by submission ID, since the headline ratios come from NVIDIA's own selection of comparison pairs.

Related Articles

benchmark

NVIDIA Vera Rubin NVL72 posts over 10x GB300 NVL72 throughput per kW on MLPerf Endpoints v1.0

Provisional MLPerf Endpoints v1.0 results show NVIDIA Vera Rubin NVL72 delivering over 10x the throughput per kilowatt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. The Vera Rubin results are pending MLCommons verification; the GB300 results are verified.

benchmark

Databricks launches public AI Index: open-weight models reach 30.9% of enterprise tokens, up from 7.3% in July

Databricks has launched the AI Index, a public resource combining usage data from 10,000+ companies on Unity Gateway with weekly coding-agent evaluations. Open-weight models grew from 7.3% of enterprise token traffic in July to 30.9% in November, and cost per task varies far more than token prices suggest.

product update

NVIDIA launches Nemotron for Science with Nemotron 3.5 Super VL and a chemistry dataset of nearly 50,000 questions

NVIDIA introduced Nemotron for Science on November 16, 2026. It pairs the open Nemotron 3.5 Super VL model with scientific evaluations, post-training recipes and LitmusBench v0.2, a chemistry dataset of nearly 50,000 questions.

research

NVIDIA recipe: RL post-training lifts Nemotron 3.5 Super VL accuracy a further 17% on drug-metabolism task

NVIDIA has published a recipe for post-training Nemotron 3.5 Super VL on scientific tasks using NeMo Gym and NeMo RL. In a drug-metabolism example, supervised fine-tuning improved accuracy by 12% and RL post-training lifted it a further 17% at a 16,384-token output budget.

Comments

Loading...