AI Research

CoreWeave claims Rubin trains 2.7x faster

CoreWeave posted first multi-rack Rubin MLPerf results, claiming 2.7x gain on DeepSeek-V3 671B.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

2 min read

CoreWeave multi-rack Vera Rubin NVL72 cluster for MLPerf training
Credit: CoreWeave Newsroom

Will Rubin still win once a job spans racks? CoreWeave says yes, reporting the first multi-rack NVIDIA Vera Rubin NVL72 results from a cloud provider in MLPerf Training v6.1, which measures how long a system takes to train a model to a fixed quality target.

The headline test is DeepSeek-V3 671B.

CoreWeave trained the 671-billion-parameter mixture-of-experts model to target quality in 17.65 minutes on 256 Rubin GPUs across multiple racks, according to the company. That run used NVIDIA Spectrum-X Ethernet networking and was the only cloud Rubin submission on that test, the company said. At the same 256-GPU count, the result is 2.7 times faster than CoreWeave's GB200 NVL72 submission, according to CoreWeave's derivation from published MLCommons results.

On a single 72-GPU Vera Rubin NVL72 rack, CoreWeave reached target on Llama 3.1 8B in 7.16 minutes, the fastest 72-GPU submission, according to the company. That is 1.9 times faster than its 13.70-minute GB200 run at the same scale, per the same derivation. Its engineering breakdown notes one rack could in principle complete eight back-to-back runs of that test in an hour, before startup and overhead. CoreWeave also reported GPT-OSS 20B in 17.18 minutes on 72 GB200 GPUs and Qwen3.5 397B reinforcement learning in 22.78 minutes on 256 GPUs using NVIDIA NeMo-RL and TensorRT-LLM, according to the company.

A new chip only pays if the fabric and operations keep every GPU busy. CoreWeave ran the tests on production clusters that customers use for frontier training, with Mission Control for fleet health and a RoCE-based Spectrum-X fabric with ConnectX-9 SuperNICs linking racks, the company said. Chen Goldberg, EVP of product and engineering at CoreWeave, said, "New silicon is only as fast as the cloud it runs on." The same MLPerf Training v6.1 round also saw Crusoe tie Nvidia at 72 GPUs on cloud.

So should planners budget Rubin as a straight 2.7x saving? Not yet, because CoreWeave notes the speedups compare its own submissions and were not verified by MLCommons. Planners can now check full results at MLCommons and watch whether rivals match multi-rack Rubin in the Closed division's Available category, which covers hardware customers can rent or buy today.

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsCoreWeaveNvidiaVera RubinMLPerfDeepSeek-V3Llama 3.1Spectrum-XMission Control

StartupHub.ai StaffDaniel SingerWritten by StartupHub.ai Staff, Editorial teamReviewed by Daniel Singer, EditorAbout

StartupHub.ai Staff, Editorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer, Editor

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

More from Daniel Singer

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.