Nvidia Vera Rubin’s MLPerf Debut: 1.94x GB300 on DeepSeek-V3

Nvidia’s Vera Rubin NVL72 trained DeepSeek-V3 671B to target quality in 4.14 minutes on 1,024 GPUs in MLPerf Training v6.1, the benchmark round MLCommons published on 16 November. The same 1,024-GPU setup on GB300 NVL72 took 8.03 minutes, so Vera Rubin was 1.94x faster, by our arithmetic from the published result logs.

Our number: the round’s fastest run, 1.38 minutes on 10,240 GB300 GPUs, used about 14,130 GPU-minutes, about 3.3x the roughly 4,240 GPU-minutes of the 1,024-GPU Vera Rubin run, by our arithmetic. CoreWeave, which announced its own Vera Rubin results at 17:00 UTC, trained the model 2.7x faster than on its GB200 NVL72 at 256 GPUs, the company says.

Business Pill · GPUs explained

An evergreen Business Engineer explainer on GPUs. It teaches the general idea only and does not discuss the MLPerf Training v6.1 results.

The key insight: As we read the MLCommons logs, Vera Rubin’s first MLPerf round shows a near-2x per-GPU gain over GB300 on the hardest test, 1.94x at 1,024 GPUs by our arithmetic, while the round’s fastest time still came from 10,240 GB300 GPUs using about 3.3x the GPU-minutes.

What the MLCommons Results Show

MLPerf Training measures “how fast systems can train models to a target quality metric”, MLCommons says on its benchmark page. The v6.1 round’s result logs are in MLCommons’ results repository, which went up at 16:51 UTC; we computed each entry’s time from those logs using the MLPerf method, and every time NVIDIA or CoreWeave also publish matches ours.

On DeepSeek-V3 671B, NVIDIA’s own entries put Vera Rubin ahead of GB300 by 1.95x at 256 GPUs (16.12 against 31.40 minutes), 1.91x at 512 GPUs (8.16 against 15.60) and 1.94x at 1,024 GPUs, by our arithmetic. At 1,024 GPUs the older GB200 NVL72 took 12.36 minutes, 2.98x Vera Rubin’s time.

NVIDIA’s technical blog puts the gain on six other benchmarks at 1.4x to 1.6x per GPU against GB300: 1.4x on Qwen3.5 397B GRPO and DLRMv4, 1.5x on Llama 2 70B LoRA and FLUX.1, and 1.6x on gpt-oss 20B and Llama 3.1 8B. The blog says Vera Rubin “delivered the fastest time to train on six of the seven benchmarks”, with GB300 at 10,240 GPUs holding the seventh.

Minutes to train DeepSeek-V3 671B in MLPerf Training v6.1, from the MLCommons result logs (computed by us per
Minutes to train DeepSeek-V3 671B in MLPerf Training v6.1, from the MLCommons result logs (computed by us per the MLPerf method; they match NVIDIA’s cited figures). Lower is faster; the comparison is not normalised for chip price or power.
DeepSeek-V3 MLPerf Training v6.1 minutes to train, GB300 NVL72 vs Vera Rubin NVL72 at 256, 512 and 1,024 GPUs
DeepSeek-V3 671B minutes to train for NVIDIA’s own GB300 NVL72 and Vera Rubin NVL72 entries, from the MLCommons MLPerf Training v6.1 result logs; times computed by us per the MLPerf method, speedups are our arithmetic.

Scale Against Efficiency

NVIDIA says the 10,240-GPU GB300 run produced “the fastest absolute result of the round: 1.38 minutes across 10,240 GPUs”. That is 3.0x faster than the 1,024-GPU Vera Rubin run with 10x the GPUs, by our arithmetic. Microsoft Azure’s 10,240-GPU GB200 entry took 2.36 minutes in the logs.

Google entered its TPU7x Ironwood chips on DeepSeek-V3. At 2,048 chips Ironwood took 5.03 minutes, 0.89 minutes longer than Vera Rubin with twice the accelerators; at 8,192 chips it took 1.70 minutes, the logs show. In GPU-minutes, the 2,048-chip Ironwood run used about 10,300, about 2.4x Vera Rubin’s, by our arithmetic.

Those GPU-minute ratios count chips, not prices or power: a TPU7x, a GB300 and a Rubin GPU are different parts, and MLPerf Training times do not include cost. NVIDIA’s blog post says Vera Rubin scaled from 256 to 1,024 GPUs “with about 97% efficiency”, which the logged times confirm, by our arithmetic.

Who Submitted What

NVIDIA says it “swept every MLPerf Training v6.1 benchmark” and that its platform “was the only one with results on every benchmark this round”. The repository shows AMD entering MI350X and MI355X systems on five benchmarks, from Llama 3.1 8B to DLRMv4, and tinycorp entering MI350X tinyboxes on two; neither has a DeepSeek-V3 or Qwen3.5 GRPO folder.

The Qwen3.5 397B GRPO test is the round’s first reinforcement-learning post-training benchmark, NVIDIA says. Vera Rubin took 10.44 minutes on it at 1,024 GPUs against 13.78 minutes for GB300, per NVIDIA’s technical blog. NVIDIA’s blog post lists submissions from 20 partner organisations on its platforms, from ASUS and Azure to Oracle and the University of Florida.

CoreWeave’s Multi-Rack Numbers

CoreWeave’s release says it was “The only cloud provider to submit multi-rack NVIDIA Vera Rubin NVL72 results”. It says it trained DeepSeek-V3 “in 17.65 minutes on 256 NVIDIA Rubin GPUs” with NVIDIA Spectrum-X Ethernet networking across multiple racks, against 48.40 minutes for its 256-GPU GB200 NVL72 entry in the logs.

NVIDIA’s own 256-GPU Vera Rubin entry took 16.12 minutes, so CoreWeave’s run took about 9% longer, by our arithmetic. On Llama 3.1 8B, CoreWeave says a single Vera Rubin rack “reached target quality in 7.16 minutes, the fastest 72-GPU submission”; NVIDIA’s own 72-GPU Vera Rubin entry took 7.38 minutes in the logs.

CoreWeave says the speedups “are derived by CoreWeave from published MLCommons results” and are not verified by MLCommons. Chen Goldberg, CoreWeave’s EVP of product and engineering, said: “New silicon is only as fast as the cloud it runs on”. The company’s backlog is in our deferred-revenue read.

Why the Timing Matters

The results came the day before Nvidia reports third-quarter fiscal 2027 results after the US close on Tuesday 17 November, with a $2.49 consensus, as our earnings preview set out. Our valuation comparison puts the stock against Broadcom and AMD, and our read of the 16 November SC26 slate covers the Vera CPU and quantum launches.

The Structural Read

As we read the results, the generational gain is steady across scale: NVIDIA’s own DeepSeek-V3 entries give 1.95x, 1.91x and 1.94x at 256, 512 and 1,024 GPUs, by our arithmetic.

As we read CoreWeave’s release, the cloud is selling proof that a rented multi-rack Vera Rubin cluster runs close to NVIDIA’s own: 17.65 minutes against NVIDIA’s 16.12 at 256 GPUs, about 9% longer by our arithmetic.

NVIDIA technical blog

“Halving time to train roughly doubles how many experiments a team can run in a quarter, and experiment throughput is what moves model quality.”

Three Implications

FOR BUYERS OF COMPUTE MLPerf Training times do not include prices, so the GPU-minute ratios we compute count chips, not cost per run.

FOR THE GOOGLE COMPARISON Ironwood at 2,048 chips took 5.03 minutes on DeepSeek-V3 in the logs, against 4.14 for 1,024 Vera Rubin GPUs.

RELATED Our earnings preview sets out the $2.49 consensus for Tuesday’s results.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Supercycle.

The framework argues: “Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”

As we read it, MLPerf v6.1 put two of those categories on the same test: NVIDIA’s Vera Rubin and GB300 against Google’s TPU7x on DeepSeek-V3, with AMD’s MI355X entered on five other benchmarks. The framework predates the results and does not discuss them.

What Is Not Established

The times are MLPerf benchmark runs on reference models, not customer training jobs. The per-GPU gains on the six other benchmarks are NVIDIA’s figures; we checked the logs for DeepSeek-V3 and Llama 3.1 8B only. NVIDIA’s technical blog also shows faster unverified reruns with newer software, which MLCommons has not reviewed.

MLPerf Training does not report prices, so the results do not show cost per training run on any of the systems. We did not contact NVIDIA, CoreWeave, Google or MLCommons for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

A Business Engineer framework mapping the nine-layer AI stack, from minerals and silicon to geopolitics.

Read the premium analysis →Get 50% off →

The Bottom Line

In MLPerf Training v6.1, Vera Rubin NVL72 trained DeepSeek-V3 in 4.14 minutes on 1,024 GPUs against 8.03 for GB300, 1.94x faster by our arithmetic from the MLCommons logs. GB300 at 10,240 GPUs set the round’s fastest time, 1.38 minutes, with about 3.3x the GPU-minutes.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Times to train come from the MLPerf Training v6.1 result logs in MLCommons’ results repository, published 16 November 2026; we computed each entry’s time with the MLPerf method (the middle of three runs, times the published scaling factor, for DeepSeek-V3; the mean of the middle eight of ten runs for Llama 3.1 8B), and every figure NVIDIA or CoreWeave also cite matches ours.

The per-benchmark gains on the other six tests, the post-training times and the partner list are from NVIDIA’s blog posts of 16 November 2026, and CoreWeave’s claims are from its release of the same day. Speedups, GPU-minutes, the scaling check and the differences between entries are our arithmetic. NVIDIA, CoreWeave, Google and MLCommons were not contacted for this piece.

Sources: MLCommons, MLPerf Training v6.1 results repository, 16 Nov 2026 · MLCommons, MLPerf Training benchmark page · NVIDIA blog, MLPerf Training and Endpoints, 16 Nov 2026 · NVIDIA technical blog, Vera Rubin in MLPerf Training v6.1, 16 Nov 2026 · CoreWeave press release, 16 Nov 2026

Scroll to Top
FourWeekMBA