Editorial illustration for NVIDIA Vera Rubin NVL72 Hits Up to 2x Blackwell Ultra in MLPerf Training
NVIDIA Vera Rubin NVL72 Doubles Blackwell Speed
NVIDIA Vera Rubin NVL72 Hits Up to 2x Blackwell Ultra in MLPerf Training
MLCommons published MLPerf Training v6.1 results this week, and NVIDIA used the round to put its next architecture on the record for the first time. The Vera Rubin NVL72 platform, making its debut submission, posted the fastest training times on six of the seven benchmarks in the suite, with gains reaching up to 2x over NVIDIA's own GB300 NVL72 baseline, a system that was already highly tuned heading into this round.
The seventh benchmark, DeepSeek-V3 671B, went to a different NVIDIA entry: a GB300 NVL72 cluster scaled to 10,240 GPUs, which claimed the fastest absolute time in the round at that benchmark. Running two platforms against each other within the same submission is not the usual move, and it raises a question about why NVIDIA chose to split its results this way rather than lead with one system across the board. NVIDIA was also the only vendor to submit results across all seven benchmarks this round, and Vera Rubin finished the full suite in its very first outing. What that scaling behavior looked like, and what it signals about the architecture's readiness, is where the numbers get specific.
MLPerf Training v6.1 marks the debut of the NVIDIA Vera Rubin architecture, and these early results demonstrate that its CUDA software stack is already highly mature. In its first appearance, the Vera Rubin NVL72 platform delivered the fastest time to train on six of the seven benchmarks, achieving up to 2x more performance than a highly optimized NVIDIA GB300 NVL72 baseline.
Why this matters
For teams planning compute budgets eighteen months out, this is the number to watch. NVIDIA is telling us, before anyone outside Santa Clara has touched the hardware, that Vera Rubin NVL72 beats Blackwell Ultra by 1.4x to 1.6x on most MLPerf Training v6.1 benchmarks and up to 2x on DeepSeek-V3. That's a vendor-run benchmark on a vendor's own platform, so treat the specific multiplier as a ceiling, not a guarantee.
Still, the breadth matters: gains across language, image generation, recommendation, and reinforcement learning suggest this isn't one cherry-picked test case. For founders scoping training runs, the real signal is software maturity, not silicon. NVIDIA is explicitly selling the claim that CUDA already runs well on hardware that's barely out the door, which is the thing that actually determines whether you can use these gains on day one or wait a year for drivers and kernels to catch up.
Researchers building agentic workflows, where iteration speed on training loops is the bottleneck, should watch for independent MLPerf submissions from other vendors before betting a roadmap on these numbers.
Common Questions Answered
How does the NVIDIA Vera Rubin NVL72 platform perform compared to the GB300 NVL72 baseline in MLPerf Training v6.1?
The Vera Rubin NVL72 platform achieved the fastest training times on six of the seven benchmarks in MLPerf Training v6.1, with performance gains reaching up to 2x over the GB300 NVL72 baseline. On most benchmarks, the Vera Rubin NVL72 beats Blackwell Ultra by 1.4x to 1.6x, demonstrating significant improvements in training efficiency across multiple workloads.
What is notable about the CUDA software stack maturity for the Vera Rubin architecture?
The CUDA software stack for Vera Rubin is already highly mature despite this being the architecture's debut submission in MLPerf Training v6.1. This early maturity is evidenced by the platform's immediate performance leadership across six of seven benchmarks, suggesting NVIDIA has invested significantly in software optimization before hardware release.
Which benchmark did a different NVIDIA entry win in MLPerf Training v6.1?
The DeepSeek-V3 671B benchmark was won by a different NVIDIA entry rather than the Vera Rubin NVL72 platform. This represents the only benchmark in the seven-benchmark suite where Vera Rubin did not achieve the fastest training time.
Why should teams planning compute budgets eighteen months out pay attention to these Vera Rubin results?
The Vera Rubin NVL72 results provide crucial performance projections for teams making long-term infrastructure investments, showing expected performance improvements of 1.4x to 1.6x on most benchmarks and up to 2x on specific workloads like DeepSeek-V3. While these are vendor-run benchmarks that should be treated as performance ceilings rather than guarantees, the breadth of gains across language and image generation tasks indicates meaningful improvements across diverse AI workloads.
Further Reading
- NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut - NVIDIA Newsroom
- NVIDIA Vera Rubin NVL72 Hits 3.7x GB300 Throughput in MLPerf v6.1 Debut - TechBeat
- MLPerf Benchmarks - NVIDIA
- MLPerf Inference v6.1: 5.7x Per-Accelerator Gains, a 512-GPU Run, and Vera Rubin's First Peer-Reviewed Numbers - Ground News