NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1
As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power. MLPerf...
First seen by Oriel on .
AI summary
NVIDIA reports that Vera Rubin NVL72 debuted in MLPerf Training v6.1 with the fastest results on six of seven benchmarks and up to 2x the performance of GB300 NVL72.
- Vera Rubin NVL72 was 1.4x to 1.6x faster than optimized GB300 NVL72 across most workloads, including language, diffusion, recommendation, and reinforcement learning.
- On DeepSeek-V3 671B, Vera Rubin trained in 4.14 minutes using 1,024 GPUs, compared with 8.03 minutes for GB300 at the same scale.
- GB300 NVL72 achieved the round’s fastest absolute DeepSeek-V3 result, completing training in 1.38 minutes across 10,240 GPUs.
- Vera Rubin maintained above 97% scaling efficiency from 256 to 1,024 GPUs, using NVLink 6, Quantum-X800 InfiniBand, or Spectrum-X Ethernet.
- Updated, unverified software reruns improved training times further, while an NVFP4 recipe delivered 8% faster pretraining and fine-tuning and 20% faster reinforcement learning on GB300.
#vera rubin #mlperf #ai training #nvlink
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| NVIDIA Developer Blog | https://developer.nvidia.com/blog/?p=125323 | https://developer.nvidia.com/blog/nvidia-vera-rubin-nvl72-achieves-up-to-2x-blackwell-ultra-performance-in-mlperf-training-v6-1/ Read discussion |