SDSignal Desk

Vera Rubin’s MLPerf debut: a training sweep nobody else contested

· Updated · Signal Desk newsroom

Analysis of reporting by NVIDIA Blog. How we work

Image: NVIDIA Blog

NVIDIA posted results on every MLPerf Training v6.1 benchmark and claims 10x-plus agentic efficiency gains in the new Endpoints test. The numbers are strong; the near-absence of rival submissions limits what “sweep” means.

Key takeaways

  • NVIDIA was the only platform with results on every MLPerf Training v6.1 benchmark, spanning LLMs, image generation, recommenders and RL post-training.
  • Vera Rubin NVL72 delivered up to 2x more performance per GPU than GB300 NVL72 in training.
  • On DeepSeek-V3, scaling from 256 to 1,024 GPUs kept about 97% efficiency, cutting time to train from 16 minutes to about four.
  • NVIDIA ran 10,240 GB300 GPUs, the largest scale this round; Microsoft Azure matched it with 10,240 GB200 GPUs.
  • In MLPerf Endpoints, a provisional Kimi K3 result showed more than 10x GB300’s throughput per watt at the same response speed.

What NVIDIA showed

Vera Rubin made its MLPerf Training debut, and NVIDIA’s summary is that it won everything. The platform was the only one with results across every v6.1 benchmark, from large language models to image generation, recommender systems and reinforcement-learning post-training. In its first outing, Vera Rubin NVL72 delivered up to twice the per-GPU performance of the GB300 NVL72, which NVIDIA describes as highly optimized after a full cycle of software tuning.

The scaling results are arguably more important for buyers than peak numbers. On the DeepSeek-V3 benchmark, Vera Rubin scaled from 256 to 1,024 GPUs at about 97% efficiency, so quadrupling hardware cut time to train from 16 minutes to roughly four. NVIDIA also ran 10,240 GB300 GPUs, the largest scale of the round, and Microsoft Azure matched it with 10,240 GB200 GPUs. Twenty partner organizations submitted on NVIDIA hardware, from Dell and HPE to CoreWeave, Google Cloud, Oracle and the University of Florida.

What “sweep” means here

A sweep is only as impressive as the field. NVIDIA’s post does not name competing accelerator submissions, and being the only platform with results on every benchmark tells us as much about who else showed up as about Vera Rubin. That is not a knock on the engineering; MLPerf remains the most standardized public comparison in the industry. It is a reason to read “leading” as leading against prior NVIDIA generations and partner systems, not necessarily against every rival chip on the market.

Generational comparisons still matter. Customers deciding between buying GB300 capacity now or waiting for Vera Rubin want to know the gap, and up to 2x per GPU in training, alongside the up to 3.7x inference gap NVIDIA claimed in September, is a meaningful answer. NVIDIA’s pitch that performance will keep climbing through software updates is plausible given its history, but it is a promise, not a result.

The agentic efficiency claim

The newer and more interesting test is MLPerf Endpoints, which measures agentic coding and enterprise workflows through API endpoints and checks that agents still complete tasks correctly. NVIDIA says Vera Rubin leads all of its agentic benchmarks, highlighted by a provisional Kimi K3 result showing more than 10x the throughput per watt of GB300 NVL72 at the same response speed.

Two cautions. The Vera Rubin Endpoints results are provisional and pending MLCommons verification, and the detailed technical write-up shows the two systems were compared at very different concurrency levels. Throughput per watt is the right metric for power-limited data centers, and NVIDIA is right that power is the binding constraint. But the 10x figure should be read as a best-case comparison until the verified results and independent tests confirm it.

Who it affects

Cloud and AI buyers
Up to 2x per-GPU training gains and near-linear scaling strengthen the case for Vera Rubin capacity.
Data center operators
Throughput per watt is the metric that decides how many agents a fixed power budget can serve.
Rival chipmakers
Absence from a full MLPerf slate leaves NVIDIA to define the public benchmark story.

What to watch

  1. MLCommons verification of the provisional Vera Rubin Endpoints results.
  2. Competing accelerator submissions in the next MLPerf rounds.
  3. Software-driven gains on Vera Rubin as systems reach customers.

FAQ

What is MLPerf Training?

A benchmark suite that measures how quickly systems train models to a target accuracy across tasks like language models, image generation and recommenders.

How much faster is Vera Rubin than GB300?

NVIDIA reports up to 2x more performance per GPU in MLPerf Training v6.1.

Are the 10x agentic results final?

No. The Vera Rubin MLPerf Endpoints results are provisional and pending MLCommons verification.

Read the original

Continue at the source.

NVIDIA Blog

Companies: NVIDIA

More on NVIDIA

All NVIDIA coverage →