Nvidia's Vera Rubin NVL72 debuted in MLPerf Training v6.1, posting up to 2x per-GPU gains over GB300 and near-linear scaling on DeepSeek-V3

Mon 16 Nov 2026 · 20:16 EST · 2 min read

In AI

Share

Nvidia’s newest AI hardware has finished its first public exam, and it topped every subject.

The Vera Rubin NVL72 rack-scale system made its debut in the MLPerf Training v6.1 benchmarks on November 16, 2026. It led all seven tests and posted the fastest training times on six of them.

The numbers behind the sweep

The headline figure is a per-GPU improvement of up to 2x over the GB300 NVL72. That baseline is no slouch, since it is the highly optimized Blackwell Ultra system Vera Rubin replaces.

The gains showed up across a demanding mix of workloads. Those included large mixture of experts (MoE) models, dense large language models (LLMs), and advanced recommendation systems.

The standout result came on the DeepSeek-V3 671B benchmark. Using 1,024 Rubin GPUs, the system finished training in 4.14 minutes.

AI, tech, and the markets they move—in one daily briefing.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

We respect your privacy. Unsubscribe anytime.

That works out to roughly double the per-GPU training speed of the GB300 at the same scale.

Why the scaling number matters most

On DeepSeek-V3, Vera Rubin reached approximately 97% efficiency when going from 256 to 1,024 GPUs.

Efficiency gains beyond training

Vera Rubin also made its first appearance in MLPerf Endpoints v1.0. There, it delivered more than 10 times the throughput per watt of the GB300 NVL72 on the Kimi K3 model under identical interactivity conditions.

CoreWeave added an independent data point. The cloud provider said its multi-rack Vera Rubin submission on DeepSeek-V3 at 256 GPUs ran 2.7 times faster than its previous GB200 NVL72 setup.

What’s under the hood

Vera Rubin keeps the familiar rack layout of 72 GPUs paired with 36 CPUs. The platform uses HBM4 memory and sixth-generation NVLink interconnects.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.