CoreWeave publishes four-rack AI training results, claiming a 2.7x speedup
The comparison holds GPU count constant against CoreWeave’s GB200 systems. Its speedup calculations and ranking claims are not verified by MLCommons.
Loading page…
The comparison holds GPU count constant against CoreWeave’s GB200 systems. Its speedup calculations and ranking claims are not verified by MLCommons.
Listen to this story
MLPerf Training v6.1 provides a cross-rack datapoint for CoreWeave’s production Vera Rubin NVL72 systems: the company says a 256-GPU DeepSeek-V3 run reached the benchmark’s quality target in 17.65 minutes. The result highlights how networking and cluster operations, alongside accelerators, shape performance when training spans multiple racks. CoreWeave’s 2.7× comparison with GB200 is its own calculation, not an MLCommons-verified ranking; NVIDIA reported a faster overall result using a much larger 10,240-GPU GB300 system.
The MLPerf round covered 153 systems and 250 results; cloud systems made up 37%, up from 29% in the previous round.
CoreWeave used NVIDIA Spectrum-X Ethernet and ConnectX-9 interfaces for traffic crossing between racks.
CoreWeave says it was the only cloud provider to submit multi-rack Vera Rubin results; that participation claim does not mean it had the fastest overall result.
A four-rack AI training job reached its benchmark quality target in 17.65 minutes on CoreWeave’s NVIDIA Vera Rubin systems, the company said November 16. Its MLPerf Training v6.1 results put 256 GPUs to work on DeepSeek-V3, offering a concrete measure of how the cloud provider’s infrastructure performs when training stretches beyond one rack.
CoreWeave’s comparable GB200 NVL72 run took 48.40 minutes at the same GPU count. The company calculates a 2.7x speedup from the published results. That calculation, along with its claims about where its submissions rank, is CoreWeave’s analysis and is not verified by MLCommons.
MLPerf Training measures how long a complete system takes to train a defined model to a specified quality target. The 17.65-minute figure is that benchmark measurement—not an unrestricted claim about the time needed for every DeepSeek-V3 training project. CoreWeave entered both Vera Rubin NVL72 and GB200 NVL72 systems in the Closed division’s Available category.
The benchmark organizer, MLCommons, released the round on November 16, 2026. It describes the suite as full-system tests that stress models, software and hardware together. This round included 153 systems and 250 results; 57 systems were cloud-based, representing 37% of the total, up from 29% in the previous round.
DeepSeek-V3’s 671-billion-parameter mixture-of-experts design routes tokens among specialized model components. That creates substantial communication demands across GPUs. In CoreWeave’s four-rack configuration, some traffic crosses the Ethernet network between racks rather than staying within the NVLink connections inside a rack. GPUs waiting for that traffic are not doing useful training work.
CoreWeave connected its racks with NVIDIA Spectrum-X Ethernet and ConnectX-9 network interfaces. It says its engineering combines networking, software, cluster operations and workload placement that accounts for how GPUs are connected. The company also monitors slow or unhealthy components, since a lagging processor or an interruption can reduce productive time across a large job.
Its provider-specific distinction is narrow: CoreWeave says it was the only cloud provider to submit multi-rack Vera Rubin results in this round. That is a participation claim, not a claim to the fastest DeepSeek-V3 result overall. NVIDIA reported that distinction for a much larger, 10,240-GPU GB300 system, which completed the benchmark in 1.38 minutes.
The rest of CoreWeave’s submission tests different demands on the system, rather than simply repeating the largest workload at a smaller scale:
MLCommons introduced the Qwen workload in this round as its first large-scale language-model post-training benchmark. It tests learning after a model’s initial training, including software-engineering tasks such as tool use. Coordinating output generation and model updates creates a different workload from pretraining alone.
For customers, CoreWeave’s practical claim is that the results used the same production infrastructure foundation available for their workloads. It says Vera Rubin NVL72 is running in production today, positioning these submissions as evidence about an operating cloud system rather than accelerator specifications alone.
Loading discussion...
Join the conversation
Explain how much weight transparency deserves when competitors haven’t submitted comparable results.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.