CoreWeave Sets New Multi-Rack AI Training Benchmark with NVIDIA Vera Rubin
Event summary
- CoreWeave achieved first multi-rack results on NVIDIA Vera Rubin NVL72 in MLPerf Training v6.1.
- Trained DeepSeek-V3 671B model in 17.65 minutes using 256 Vera Rubin GPUs across multiple racks.
- Vera Rubin performance was 2.7x faster than GB200 NVL72 for same model at 256 GPUs.
- Fastest 72-GPU submission on Llama 3.1 8B, completing in 7.16 minutes.
- Benchmarks run on production clusters using CoreWeave Mission Control.
The big picture
CoreWeave's benchmark results demonstrate the practical scaling advantages of NVIDIA's Vera Rubin architecture in cloud environments. This positions CoreWeave as a leader in high-performance AI infrastructure as model sizes continue to grow. The results validate CoreWeave's integrated approach to AI cloud services, combining cutting-edge hardware with production-grade orchestration systems.
What we're watching
- Performance Scaling
- How CoreWeave's multi-rack performance will translate to real-world AI training workloads.
- Competitive Positioning
- Whether CoreWeave can maintain its benchmark leadership as competitors adopt Vera Rubin.
- Customer Adoption
- The pace at which AI developers migrate to Vera Rubin-based cloud infrastructure.
Related topics
