Hot AI News

Vera Rubin cuts DeepSeek-V3 training to 4 minutes in its MLPerf debut

Nvidia says Vera Rubin NVL72 delivers up to 2x the per-GPU training performance of GB300 and, in a provisional result, more than 10x the agentic throughput per watt.

Vera Rubin cuts DeepSeek-V3 training to 4 minutes in its MLPerf debut
Image: Nvidia

Nvidia says its Vera Rubin platform, in its first MLPerf appearance, trained models up to 2x faster per GPU than GB300 and swept every MLPerf Training v6.1 benchmark. A provisional agentic-inference result claims more than 10x the throughput per watt.

Training: Twice the per-GPU speed, with near-linear scaling

According to Nvidia, it was the only platform to submit results on every benchmark in the round. They cover large language models, image generation, recommenders and reinforcement learning post-training. Vera Rubin NVL72 delivered up to 2x more performance per GPU than the GB300 NVL72 system, which Nvidia calls highly optimized.

DeepSeek-V3 time to train, 1,024 GPUs
Image: Nvidia, chart by Aiholics

Nvidia’s developer blog breaks the gain down by workload. The 2x figure applies to DeepSeek-V3 671B on 1,024 GPUs. Other benchmarks show smaller gains: 1.4x for the DLRMv4 recommender and for Qwen3.5 397B reinforcement learning, 1.5x for Llama 2 70B LoRA and FLUX.1, and 1.6x for Llama 3.1 8B and gpt-oss 20B.

On DeepSeek-V3, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at about 97% efficiency. That cut training time from 16 minutes to about four. Nvidia’s technical post lists 4.14 minutes for Vera Rubin on 1,024 GPUs, against 8.03 minutes for GB300 and 12.36 minutes for GB200.

Chart of MLPerf Training v6.1 results for Nvidia Vera Rubin NVL72
Image: Nvidia

Four times the GPUs cut DeepSeek-V3 training time from 16 minutes to about four.

The largest scale in the round was 10,240 GPUs. Nvidia ran it on GB300 systems, and Microsoft Azure matched it with GB200 systems. Nvidia’s technical post says the GB300 run set the round’s fastest absolute DeepSeek-V3 time, at 1.38 minutes. Submissions came from 20 organizations, including Dell, HPE, Google Cloud, CoreWeave, Oracle and Supermicro.

Chart of MLPerf Training v6.1 end-to-end results across benchmarks
Image: Nvidia
Advertisements

Endpoints: A 10x power-efficiency claim that is still provisional

MLPerf Endpoints is a newer benchmark. It runs agentic coding and enterprise workflows through API endpoints and checks that the agents still complete their tasks correctly. In its first submission, Nvidia reports that Vera Rubin NVL72 delivers more than 10x the throughput per kilowatt of GB300 NVL72 on Kimi K3, at a shared minimum of 50 tokens per second per user.

The Kimi K3 numbers come from different setups. The Vera Rubin run used 32 accelerators at 1,024 concurrent queries and drew 100.89 kW, producing 41,414.2 tokens per second. The GB300 run used 48 accelerators at 48 queries and drew 60 kW, producing 2,210.4 tokens per second. Nvidia’s technical post gives other ratios: about 9x per kilowatt on DeepSeek-V4.1-Flash and about 5x on Qwen3.6-35B-A3B.

Nvidia says the Vera Rubin results are pending MLCommons verification and could change after peer review. Only the GB300 results are verified.

What to watch before treating the numbers as settled

The comparisons are all Nvidia against Nvidia. Neither Nvidia post names a competing accelerator in these results, so they show generational progress rather than a lead over rivals. The “sweep” is also a statement about participation: Nvidia says it was the only platform with results on every benchmark. Its developer blog says about 80% of closed-division submissions ran on Nvidia hardware.

Three things would confirm or weaken the claims:

  • MLCommons verification of the provisional Endpoints results, which could move the 10x figure.
  • Independent workloads. Nvidia also cites SemiAnalysis AgentX and Artificial Analysis AgentPerf, where it says Vera Rubin leads, but gives no figures for them.
  • The software gains Nvidia promises. Its developer blog shows unverified reruns on a newer software stack, with no MLCommons review. Llama 3.1 8B improved from 7.162 to 6.140 minutes on Vera Rubin, and gpt-oss 20B from 11.213 to 8.080 minutes. Official results in later rounds should show whether that holds.

The blog’s tables also have an inconsistency. The Qwen3.5 reinforcement learning row lists 64 GPUs in the table but 1,024 in the text, with different submission IDs. That makes the exact configuration behind the 1.4x figure unclear.

Availability is not covered. Nvidia gives no shipping dates or pricing for Vera Rubin NVL72 in these posts, and the Endpoints entries sit in MLPerf’s Available category. This is separate from the Rubin HPC GPU that our earlier report said would not ship until 2028.

For buyers, the Endpoints figure matters because AI data centers are limited by power. Tokens per watt is the number that decides how many agents a fixed power budget can serve. A gain that large would change capacity planning, but only if the verified result lands close to the provisional one.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like