The AI Sift is part of you-do-nothing

← Back to The Latest

Nvidia swept all seven MLPerf Training tests with Vera Rubin, running 2x faster per GPU

Nvidia swept all 7 MLPerf Training tests on November 16 with Vera Rubin, up to 2x faster per GPU than the GB300 system it replaces.

The decision it puts on your desk

Re-price your 2027 compute budget within 7 days, using power as the unit. Lock orders by January 31.

Nvidia swept all seven MLPerf Training tests on November 16, with Vera Rubin running up to 2x faster per GPU.

Vera Rubin racks in a data center (Source: Nvidia)
Vera Rubin racks in a data center (Source: Nvidia)

Training results

Nvidia was the only vendor with a result on every benchmark in the round. The tests span language models, image generation, recommender systems and post-training. Independent coverage counted seven tests led and the fastest training time on six of them.

"These results are just the beginning," Nvidia said in a post on Monday. Vera Rubin was up to 2x faster per GPU than the GB300 (the older system). The company expects the same machines to post higher numbers as software work continues.

What the sweep actually shows
What the sweep actually shows

Scaling and power

On the DeepSeek-V3 run, Vera Rubin finished in 4.14 minutes using 1,024 GPUs. Going from 256 to 1,024 chips held about 97 percent efficiency. Nvidia also ran the test on 10,240 GB300 GPUs, the largest scale in the round.

Nvidia frames power as the real limit in a data center. "Because AI data centers are limited by power, the more tokens a system can generate each second for every watt of power, the more agents it can serve and the more value it creates," Nvidia said. Power is what a buyer pays for each month.

Verification gaps

The biggest number in the release carries a caveat. The 10x per watt figure on Kimi K3 is provisional, and MLCommons has not checked it. The GB300 run beside it did pass that check.

The two systems also differ in size: 32 accelerators against 48. Both were held to the same floor of 50 tokens per second per user. CoreWeave said a 256 chip Vera Rubin run beat its previous system by 2.7 times.

The 10x number, checked against the published figures
The 10x number, checked against the published figures

Competitor response

Twenty organizations submitted to the round, among them Google Cloud, Oracle and CoreWeave. Azure matched Nvidia's largest run with 10,240 GB200 (its previous generation) chips. The comparison base is Nvidia's own previous system.

The 2x number therefore measures a step rather than a win over a rival. CoreWeave published its own run as an independent data point beside the vendor claims. Nvidia also points at SemiAnalysis and Artificial Analysis numbers that sit outside MLPerf.

Market context

The results land at the supercomputing conference running this week. AI shares are weak this week, with Anthropic's listing expectations slipping on the same days. Buyers setting 2027 capacity contracts are reading these numbers right now.