NVIDIA quietly cuts Vera Rubin specs, HBM4 bandwidth and NVLink 6 speeds now lower
NVIDIA has revised the technical specifications for its Vera Rubin AI platform. The updated data shows a reduction in advertised HBM4 memory bandwidth and sixth-generation NVLink 6 speeds. These changes affect individual GPUs, the Vera Rubin NVL72 rack, and the HGX Rubin NVL8 platforms.
Reductions in NVL72 and HGX Platforms
The revisions significantly impact the Vera Rubin NVL72 specifications. While NVIDIA previously announced a HBM4 bandwidth of 22 TB/s per GPU, the updated figures show 19.2 TB/s, a 12.7% decrease. The memory capacity remains at 288 GB per GPU. Additionally, NVLink 6 bandwidth has been lowered from 3.6 TB/s to 3 TB/s per GPU, representing a 16.7% drop.
For the full Vera Rubin NVL72 rack, which integrates 72 Rubin GPUs and 36 Vera CPUs, the aggregate HBM4 bandwidth has been adjusted from 1,580 TB/s to 1,400 TB/s. The total NVLink bandwidth also dropped from 260 TB/s to 216 TB/s. Interestingly, scale-out networking bandwidth actually increased from 28.8 TB/s to 32.4 TB/s, while the total HBM4 capacity remains at 20.7 TB.
The HGX Rubin NVL8 eight-GPU platform saw even steeper reductions. Its aggregate HBM4 bandwidth fell by 26%, dropping from 176 TB/s to 130 TB/s, and the total memory capacity decreased from 2.3 TB to 2 TB. NVIDIA also lowered training performance metrics: from 280 to 270 PFLOPS for NVFP4 and from 140 to 130 PFLOPS for FP8/FP6.
Documentation Discrepancies
There is currently some ambiguity in the official documentation. While the HGX Rubin NVL8 product page lists the reduced specs, some individual GPU tables and the DGX Rubin NVL8 page still advertise the original 2.3 TB capacity and 176 TB/s bandwidth. It remains unclear whether these variations represent different product configurations or if the documentation simply hasn't been fully synchronized.
Гика
