CoreWeave, NVIDIA and IBM Set MLPerf Record With 2,496-GPU Blackwell Cluster

The largest Blackwell submission to date

The joint entry ran on 2,496 NVIDIA Blackwell GPUs spread across 39 racks, each with 64 active GPUs, hosted on CoreWeave's infrastructure. According to CoreWeave, it was over 34x larger than the only other GB200 Grace Blackwell submission from a cloud provider, a single rack with 72 GPUs. It was also 4.8x larger than the next largest entry, NVIDIA's own submission with 512 Blackwell GPUs.

CoreWeave says the submission is equivalent in computing power to more than 5,000 NVIDIA H100 GPUs. That figure assumes a training speedup of 2.15x on the Llama 3.1 405B benchmark over the fastest H100 submission. It is the largest training benchmark CoreWeave has delivered, ahead of its earlier MLPerf Training v3.1 submission with 3,584 H100 GPUs.

The run took place on the Carina cluster, which CoreWeave built and delivered for IBM, with technical expertise from NVIDIA for the benchmarking. CoreWeave describes Carina as currently one of the largest operational Blackwell clusters in the world.

Results on Llama 3.1 405B

MLPerf is a benchmark suite developed by MLCommons that measures how quickly a system can train leading open-source models to a defined quality target. In version 5.0, MLCommons added Meta's Llama 3.1 405B, which the source describes as the largest and most complex foundational model in the suite.

CoreWeave reports these times for Llama 3.1 405B on its Blackwell configurations:

• 2,496 GPUs: 27.33 minutes, 2.1x faster than NVIDIA's larger 2,560 H100 GPU submission
• 2,048 GPUs: 32.6 minutes, 2.15x faster than NVIDIA's 2,048 H100 GPU submission
• 1,536 GPUs: 42.46 minutes, 2.85x faster than CoreWeave's own 512 Blackwell GPU submission
• 1,024 GPUs: 62.1 minutes, 2x faster than the top 1,024 H200 GPU submission
• 512 GPUs: 121.1 minutes, 2x faster than the top 512 H200 GPU submission

The company says its cluster posted the fastest training times across five cluster configurations when comparing similar-sized clusters. The largest submission reached 91% scaling efficiency from 512 to 2,496 GPUs. CoreWeave also expects further gains from ongoing software and system-level optimizations.

Density, software and earlier records

CoreWeave highlights the space savings of the GB200 NVL72 system. The 2,496-GPU run needed 39 racks, while an equivalent H100 setup, assuming comparable performance and 32 GPUs per rack, would have needed around 156 racks. The company calls this a 4x reduction in physical space and points to energy savings from higher-density configurations.

On the software side, the company says CoreWeave Kubernetes Service and Mission Control provide deep observability, bare-metal performance isolation and proactive infrastructure management, maintaining up to 96% goodput for large-scale training. The benchmarks ran on SUNK, its Slurm-on-Kubernetes solution, which schedules jobs within GB200 NVL72 systems for rack-level efficiency over NVIDIA NVLink.

The company also recalled earlier results. In MLPerf Inference v5.0, it was the first cloud provider to submit with GB200 Grace Blackwell Superchips, reaching over 200 tokens per second per chip on Llama 3.1 405B, a 2.86x per-chip increase over NVIDIA H200 GPUs. In MLPerf Training v3.0, it completed the GPT-3 175B benchmark in under 11 minutes using more than 3,500 H100 GPUs.

What it means for customers

CoreWeave presents the results as evidence for customers building, fine-tuning or deploying AI applications on its platform. The figures are the company's own claims, though the source states that MLCommons Association verified the results. For IBM, the run shows a large Blackwell cluster operating reliably, with CoreWeave saying it handled multiple training runs above 1,000 GPUs. Buyers weighing infrastructure can read the numbers as a comparison of speed, scale and rack density between Blackwell and Hopper generations.