ExtremeTech는 Nvidia가 Vera Rubin의 성능이 당초 주장했던 것만큼 강력하지 않다고 인정했다고 보도했다.

Nvidia has quietly revised the specifications of its next-generation Vera Rubin GPUs, affecting individual chips, its NVL72 rack, and the HGX Rubin NVL8 platform. The new spec table lists the GPUs with memory bandwidth reduced by almost 13% and NVLink 6 bandwidth reduced by nearly 17%.

Nvidia debuted the Vera Rubin GPU at CES 2026. At the time, it showcased the NVL72 AI server system for data centers, highlighting its 288GB of HBM4 memory and up to 22TB/s of bandwidth per GPU. The new specifications list that bandwidth at 19.2TB/s, while memory capacity remains the same.

The HGX Rubin NVL8 system integrates eight Rubin GPUs with NVLink 6 networking connections. Its memory bandwidth is now listed at 130TB/s, down from 176TB/s, a 26% drop from the original figure. Memory capacity also falls from 2.3TB to 2TB, and AI training performance is reduced. NVFP4 training drops from 280 PFLOPS to 270 PFLOPS, and FP8/FP6 training drops from 140 PFLOPS to 130 PFLOPS.

On the largest system, the NVL72 Vera Rubin rack, which integrates 72 GPUs with 36 Vera CPUs, memory bandwidth is now listed at 1,400TB/s rather than 1,580TB/s. Memory capacity remains the same.

Nvidia's listed NVLink bandwidth is also down across all systems. On a per-GPU basis, it falls from 3.6TB/s to 3TB/s. For the HGX NVL8 system, it drops from 28.8TB/s to 24TB/s, and for the NVL72, from 260TB/s to 216TB/s.

The only positive revision is scale-out networking bandwidth, which increased from 28.8TB/s to 32.4TB/s.

The changes appeared on Nvidia's website in early September. The company has not announced them or explained why they were made.

None of this is likely to affect demand for Vera Rubin. However, it matters alongside growing competition from companies like AMD and Chinese alternatives in the international market. It is also a reminder to view initial product announcements with a healthy dose of skepticism.