Nvidia has revised the technical specifications for its Vera Rubin AI platform, reducing memory bandwidth and interface speeds. The changes affect individual Rubin GPUs, the Vera Rubin NVL72 rack system, and the HGX Rubin NVL8 platform.
For the Vera Rubin NVL72 system, the HBM4 memory bandwidth per GPU was lowered from 22 TB/s to 19.2 TB/s. The NVLink 6 interface speed per GPU dropped from 3.6 TB/s to 3 TB/s. Consequently, the aggregate bandwidth for the 72-GPU system decreased to 1400 TB/s for memory and 216 TB/s for NVLink.
The HGX Rubin NVL8 platform saw more significant adjustments. Total HBM4 memory capacity was reduced from 2.3 TB to 2 TB, while memory bandwidth fell from 176 TB/s to 130 TB/s. NVLink Switch bandwidth decreased from 28.8 TB/s to 24 TB/s.
Performance metrics for model training were also adjusted. For the NVFP4 format, performance dropped from 280 PFLOPS to 270 PFLOPS, and for FP8/FP6 formats, it fell from 140 PFLOPS to 130 PFLOPS.
Discrepancies remain in Nvidia's documentation, as some product pages still display the original, higher specifications. According to Wccftech, the updates for the NVL72 system were implemented between September 1 and September 11.
