Nvidia Vera Rubin Leads MLPerf Training Benchmarks

The rack-scale NVL72 system debuted in MLPerf Training v6.1, leading all seven tests in the benchmark suite.

Updated on Nov. 17, 2026 in Artificial Intelligence

Bold flat-color economist-style illustration of a dark server rack with copper coolant pipes on a deep navy and cream field.
Nvidia's Vera Rubin NVL72 rack-scale system led all seven tests in MLPerf Training v6.1, posting up to a 2x per-GPU gain over the GB300 NVL72. AI Illustration. Upload story photo >

Nvidia's Vera Rubin NVL72 rack-scale system debuted in MLPerf Training v6.1 on November 16, 2026, leading all seven tests and posting the fastest training times on six of them. The system achieved up to a 2x per-GPU improvement over the GB300 NVL72.

Why it matters

The results offer a benchmark comparison of training performance for large AI models, including mixture of experts and dense large language models. Faster training times and per-GPU gains bear directly on the cost and speed of building advanced AI systems worldwide.

Each Vera Rubin rack pairs 72 GPUs with 36 CPUs, HBM4 memory and sixth-generation NVLink interconnects. On the DeepSeek-V3 671B benchmark, 1,024 Rubin GPUs finished training in 4.14 minutes, with roughly 97% scaling efficiency from 256 to 1,024 GPUs.

The players

Nvidia

Nvidia is a American semiconductor company best known for designing the GPUs that power most modern AI training and inference systems.

CoreWeave

CoreWeave is a cloud infrastructure provider specializing in GPU-accelerated computing for AI workloads.

The details

The system was benchmarked across large mixture of experts models, dense large language models and advanced recommendation systems. Vera Rubin also made its first appearance in MLPerf Endpoints v1.0, where it delivered more than 10 times the throughput per watt of the GB300 NVL72 on the Kimi K3 model.

Timeline

  1. November 16, 2026: Vera Rubin NVL72 debuts in MLPerf Training v6.1 benchmarks.

The Tech Race

The Vera Rubin NVL72 succeeds the GB300 NVL72 generation, with the MLPerf Training v6.1 results showing up to a 2x per-GPU gain over its predecessor. The results intensify competition among chipmakers to lead the standardized MLPerf benchmark suite.

Faster training per GPU can shorten the time it takes developers to build and update large AI models. Gains in throughput per watt also bear on the energy and running costs of AI services that users rely on.

The takeaway

Benchmark leadership in training speed and efficiency signals where AI infrastructure spending is heading. Readers tracking AI tools should watch how these per-GPU gains translate into cheaper and faster model updates.

Further reading

For more on the chips and systems driving this progress, explore our Artificial Intelligence coverage.