Today ainotis Join

My notis

ResearchPublished Top storyAll news from that day

Nvidia says Vera Rubin NVL72 trained DeepSeek-V3 in 4.14 minutes on 1,024 GPUs

In Nvidia's own MLPerf Training v6.1 account, GB300 NVL72 took 8.03 minutes on the same GPUs; Vera Rubin's drop from 16 minutes on 256 GPUs to about four came from four times the GPUs.

Share

Check our sources · 11 facts from 3 sources
Two facing rows of dark data-centre racks with lit, cream-coloured server trays, reflected in a black floor.Image: NVIDIA
source · The company's own image of its product, used to report on it

Key points

  1. On the same 1,024 GPUs, Nvidia's developer blog says Vera Rubin NVL72 took 4.14 minutes to train DeepSeek-V3 671B and GB300 NVL72 took 8.03 minutes.
  2. The fastest absolute DeepSeek-V3 time was GB300 NVL72 at 1.38 minutes on 10,240 GPUs, Nvidia says; MLCommons published the round on 16 November, with 153 systems.
  3. Nvidia's provisional Kimi K3 claim of more than 10x the throughput per watt of GB300 NVL72 is pending MLCommons verification and subject to change.

What happened

Nvidia says Vera Rubin NVL72 made its MLPerf Training debut in Training v6.1 with up to twice the performance per GPU of GB300 NVL72, and calls the results early. MLCommons published the Training v6.1 results on 16 November: 24 submitting organisations, 153 systems and 250 test results, and it says the submissions included NVIDIA Rubin-generation GPUs for the first time.

On DeepSeek-V3, Nvidia says Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97 percent efficiency, and that quadrupling the GPUs cut time to train from 16 minutes to about four. That drop comes from using four times the GPUs, not from a newer chip.

On the same 1,024 GPUs, Nvidia's developer blog says Vera Rubin NVL72 trained DeepSeek-V3 in 4.14 minutes against 8.03 minutes for GB300 NVL72. It says the fastest absolute time was GB300 NVL72 at 1.38 minutes on 10,240 GPUs, and that Microsoft Azure matched that scale with 10,240 GB200 GPUs.

The blog says Vera Rubin NVL72 had the fastest time to train on six of the seven benchmarks against an optimized GB300 NVL72 baseline, and cites three MLCommons entries (6.1-0045, 6.1-0030 and 6.1-0116) for its 1,024-GPU comparison without saying which entry is which.

Separately, in the first MLPerf Endpoints submission, a provisional result for the Kimi K3 model, Nvidia claims more than 10 times the throughput per watt of GB300 NVL72 at the same response speed.

Its footnote to that result says the GB300 results were verified by MLCommons Association but the provisional Vera Rubin results are pending verification and subject to change. Nvidia describes Vera Rubin NVL72 as generally available through cloud partners. The post compares Nvidia's own systems only.

What it means for you

Our view

This is Nvidia's account of its own systems: the post compares only its GB300 and GB200 hardware, so it says nothing about other vendors' chips. The drop from 16 minutes to about four reflects a larger cluster, which makes the like-for-like figure the one on 1,024 GPUs.

Nvidia calls the results early, and the Kimi K3 throughput figure is provisional until MLCommons verifies it. If you compare cloud offers for model training, ask each provider for time to train at your own GPU count, not at the largest published scale.

It adds no new facts.

Share this story

Your reaction

We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.

Check our sources

Every sentence above is checked against these 3 sources.

1 NVIDIA MLPerf Training v6.1 and MLPerf Endpoints resultsNVIDIA · 16 Nov 2026 · 6 facts Open the source archived copy
  1. Nvidia says Vera Rubin NVL72, in its MLPerf Training v6.1 debut, delivered up to 2x more performance per GPU than GB300 NVL72; the post calls these early results. the company says

    Toned down to what the source says
    In its debut, Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system — early results that show the pace of generation-over-generation innovation.
    cite
  2. Nvidia says that on DeepSeek-V3, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency, and that quadrupling the GPUs cut time to train from 16 minutes to about four. the company says

    On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs with about 97% efficiency: quadrupling the GPUs cut time to train from 16 minutes to about four.
    cite
  3. Nvidia says it ran DeepSeek-V3 on 10,240 GB300 GPUs, the largest scale in the round, and that Microsoft Azure matched that scale with 10,240 GB200 GPUs. the company says

    NVIDIA also ran the benchmark on 10,240 GPUs with GB300 systems, representing the largest scale in this round. Microsoft Azure matched that scale with 10,240 NVIDIA GPUs with GB200 systems.
    cite
  4. Nvidia says that for Kimi K3, in a provisional result in the first MLPerf Endpoints submission, Vera Rubin NVL72 delivers more than 10x the throughput per watt of GB300 NVL72 at the same response speed. the company says

    in a provisional result for Kimi K3, a large open model built for agentic workflows, Vera Rubin NVL72 systems deliver more than 10x the throughput per watt of NVIDIA GB300 NVL72 at the same response speed.
    cite
  5. Nvidia's footnote to its Kimi K3 MLPerf Endpoints result says the GB300 results were verified by MLCommons Association and that the provisional Vera Rubin results are pending MLCommons Association verification and subject to change pending peer review. the company says

    Toned down to what the source says
    MLPerf® Endpoints v1.0, Kimi K3, Standardized division, Available category, published November 9, 2026. ... GB300 results verified by MLCommons Association. For the provisional Vera Rubin results: pending MLCommons Association verification, results are subject to change pending peer review.
    cite
  6. Nvidia's post makes no comparison with other vendors' hardware; its comparisons are against Nvidia's own GB300 and GB200 systems. the company says

    Vera Rubin NVL72 delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system
    cite
2 MLPerf Training v6.1 Results: Record ParticipationMLCommons · 16 Nov 2026 · 2 facts Open the source archived copy
  1. MLCommons published the MLPerf Training v6.1 results on 16 November 2026: 24 submitting organisations, 153 systems and 250 test results, with the largest system using 10,240 accelerators.

    November 16, 2026 ... Participants in this round submitted 153 systems, yielding 250 test results. ... Twelve systems used 1,000 or more accelerators, with the largest system using 10,240 accelerators. ... The MLPerf Training v6.1 round includes performance results from 24 submitting organizations
    cite
  2. MLCommons says the v6.1 submissions included, for the first time, NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood); its page does not give their availability category.

    Systems utilized NVIDIA, AMD and Google accelerator hardware, including for the first time NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood).
    cite
3 NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1NVIDIA · 16 Nov 2026 · 3 facts Open the source archived copy
  1. Nvidia's developer blog says Vera Rubin NVL72 trained DeepSeek-V3 671B in 4.14 minutes on 1,024 GPUs, against 8.03 minutes for GB300 NVL72 on the same 1,024 GPUs, and that the fastest absolute DeepSeek-V3 time was GB300 NVL72 at 1.38 minutes on 10,240 GPUs. the company says

    Vera Rubin NVL72 trained the benchmark in 4.14 minutes using 1,024 GPUs across 16 racks. GB300 NVL72 ran the identical 1,024-GPU configuration in 8.03 minutes ... At a larger scale, GB300 NVL72 produced the fastest absolute result of the round: 1.38 minutes across 10,240 GPUs.
    cite
  2. Nvidia's developer blog says Vera Rubin NVL72 posted the fastest time to train on six of the seven MLPerf Training v6.1 benchmarks, against an optimized GB300 NVL72 baseline, and cites MLCommons entries 6.1-0045, 6.1-0030 and 6.1-0116 for its 1,024-GPU DeepSeek-V3 comparison. the company says

    Toned down to what the source says
    debuted in MLPerf Training v6.1 and delivered the fastest time to train on six of seven benchmarks, achieving up to 2x more performance per GPU than an optimized GB300 NVL72 baseline. ... MLPerfTraining v6.1 results retrieved from ... on Nov 16, 2026, from the following entries: 6.1-0045, 6.1-0030 and 6.1-0116.
    cite
  3. Nvidia's developer blog describes Vera Rubin NVL72 as generally available through cloud infrastructure partners; it says reruns with a newer software stack for Llama 3.1 8B, gpt-oss 20B and DLRMv4 are not MLCommons-verified and not part of the official results. the company says

    Vera Rubin NVL72 systems are generally available through cloud infrastructure partners worldwide. ... These are unverified results rather than MLCommons submissions
    cite

Topics

The morning email

On the mornings we publish: the three top stories and up to four short ones. Free.

We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep