Nvidia has quietly reduced the memory and interconnect bandwidth specifications for its Vera Rubin GPUs without issuing a formal announcement. The per-GPU HBM4 bandwidth has been lowered from 22 TB/s to 19.2 TB/s, while NVLink 6 bandwidth has dropped from 3.6 TB/s to 3 TB/s. The eight-GPU HGX Rubin NVL8 configuration was hit hardest, losing 26% of its memory bandwidth, falling from 2.3 TB to 2 TB of HBM4 capacity.
The changes were first spotted by IT Home, which compared Nvidia's October documentation with a June 2026 version. The edits appear to have been made between September 1 and September 11. Nvidia opted to update the product pages silently, without a press release or explanation. The timing is notable, as the adjustments came shortly after Nvidia reported that Vera Rubin racks were already operational at partners like CoreWeave, Google Cloud, and Microsoft Azure.
The downgrade raises questions about whether Nvidia is scaling back its targets to ensure timely delivery of hardware to customers. The revised figures suggest that the company may have had to settle for lower bandwidth due to challenges faced by memory suppliers like SK hynix and Samsung. AMD's Instinct MI455X now leads in per-GPU memory bandwidth, while Helios surpasses Rubin in scale-up bandwidth, at least on paper.
Despite the reductions, Nvidia's Vera Rubin remains a powerful GPU, offering about 2.4 times the bandwidth of a Blackwell GPU. However, the company has not clarified whether it has recalculated performance projections based on the lower specifications. The marketing claims, such as up to 10x more tokens per megawatt, remain unchanged, though they are subject to revision.