Skip to main content

Nvidia leads seven AI training benchmarks with Vera Rubin results

Crypto Briefing
17 lis 2026, 01:21Crypto
Oryginał
1 min read

Nvidia's Vera Rubin NVL72 rack-scale system topped all seven MLPerf Training v6.1 benchmarks on November 16, 2026, posting up to 2x per-GPU gains over its predecessor GB300.

The Vera Rubin NVL72 system completed its first public benchmark evaluation, achieving the fastest training times in six of the seven tests. The hardware demonstrated particularly strong performance on large mixture of experts (MoE) models, dense large language models (LLMs), and advanced recommendation systems.

Performance gains

The system showed up to 2x per-GPU improvement over the GB300 NVL72, which it replaces. On the DeepSeek-V3 671B benchmark, the Vera Rubin completed training in just 4.14 minutes using 1,024 GPUs, roughly double the per-GPU training speed of the GB300 at the same scale.

Scaling efficiency

The system achieved approximately 97% efficiency when scaling from 256 to 1,024 GPUs on DeepSeek-V3, demonstrating near-linear performance gains. CoreWeave independently verified these results, reporting that its multi-rack Vera Rubin setup ran 2.7 times faster than its previous GB200 NVL72 configuration.

Technical specifications

Vera Rubin maintains the familiar rack layout of 72 GPUs paired with 36 CPUs, utilizing HBM4 memory and sixth-generation NVLink interconnects. In MLPerf Endpoints v1.0 testing, it delivered more than 10 times the throughput per watt of the GB300 NVL72 on the Kimi K3 model under identical conditions.

Z.ai
Article written by AI
Z.ai GLM-4.5-FlashFree
Free AI articles150 of 150
No limit after signing in with Telegram