The Pulse
CoreWeave reports multi-rack Vera Rubin AI training results
CoreWeave says its multi-rack Vera Rubin system reached a DeepSeek-V3 training target in 17.65 minutes. Its comparison with a GB200 system at the same GPU count was not verified by MLCommons.

CoreWeave says it trained DeepSeek-V3 671B to a benchmark quality target in 17.65 minutes using 256 Nvidia Rubin GPUs spread across four racks. The company reported the result on November 16 as part of the MLPerf Training v6.1 benchmark round.
The Vera Rubin run was 2.7 times faster than CoreWeave’s GB200 NVL72 submission at the same GPU count, which took 48.40 minutes. CoreWeave derived that comparison from published MLCommons results; the comparison was not verified by the MLCommons Association.
Measuring time to quality
MLPerf Training measures how long a system takes to train a specified model to a defined quality target. CoreWeave submitted Vera Rubin NVL72 and GB200 NVL72 results in the Closed division’s Available category, which compares systems using the same reference model and hardware customers can buy or rent.
The company says it was the only cloud provider to submit multi-rack Vera Rubin results in this round. That claim is based on its analysis of the published Closed division results at the same accelerator count and was not verified by MLCommons.
Its submissions covered models ranging from 8 billion to 671 billion parameters and systems ranging from a single 72-GPU rack to 256 GPUs across four racks. The Vera Rubin results covered DeepSeek-V3 and Llama 3.1 pretraining; the company also submitted GB200 results for those workloads, GPT-OSS 20B pretraining and Qwen3.5 397B reinforcement learning.
Single-rack and GB200 results
In the smaller Llama 3.1 8B test, CoreWeave says a single Vera Rubin NVL72 rack reached the quality target in 7.16 minutes. It described that as the fastest 72-GPU submission in the round and 1.9 times faster than its GB200 result of 13.70 minutes at the same GPU count.
Both platforms used NVFP4 GEMMs, FP8 attention and full Nvidia CUDA graphs for that test. The company’s speedup calculations compare its own submissions rather than results from different providers.
For Qwen3.5 397B reinforcement learning, CoreWeave reported a 22.78-minute result on 256 GB200 GPUs across four racks. The run used 128 GPUs for training and 128 for generation, with Nvidia NeMo and TensorRT-LLM.
The company also reported training GPT-OSS 20B in 17.18 minutes on a single 72-GPU GB200 rack. The Qwen workload is a post-training test, while the GPT-OSS result is a pretraining test.
Connecting the racks
DeepSeek-V3 is a mixture-of-experts model that routes tokens between experts at each mixture-of-experts layer. In a four-rack run, some of that traffic crosses the Spectrum-X Ethernet network between racks rather than staying within the Nvidia NVLink domain.
For the multi-rack runs, CoreWeave connected racks using a RoCE-based Nvidia Spectrum-X Ethernet fabric, 102.4 Tb/s Spectrum-6 networking and ConnectX-9 SuperNICs. The company says it ran the Vera Rubin benchmarks on the same production infrastructure foundation available to customers.
CoreWeave Mission Control coordinates workload placement and cluster operations, while tools called Racky and Valvey manage racks and cooling. Its orchestration accounts for GPU and network locality, and fleet monitoring and lifecycle automation identify slow or unhealthy components.
Sources
Companies in this story
Related stories
- NVIDIA reports Vera Rubin gains in MLPerf Training debut
- Nvidia claims Vera Rubin inference lead in provisional MLPerf results
- CoreWeave cites study showing up to 65% lower GB200 NVL72 cloud costs
- CoreWeave expects about $9 billion in customer prepayments and deposits
- Google reports Ironwood TPU results in MLPerf training tests