Nvidia has quietly lowered the memory and interconnect bandwidth specifications published for its Vera Rubin architecture, cutting per-GPU HBM4 bandwidth by roughly 13 percent and NVLink bandwidth by about 17 percent without issuing a formal press release. The specification adjustments first surfaced after researchers compared recent documentation with snapshots of the manufacturer's product pages, revealing that the revised figures went unnoticed for at least six weeks.
Under the updated specifications, the Vera Rubin NVL72 product page now rates each GPU at 19.2 TB/s of HBM4 bandwidth, down from the previously advertised 22 TB/s. Similarly, NVLink 6 bandwidth has dropped from 3.6 TB/s down to 3 TB/s. While headline figures such as memory capacity and sparse NVFP4 inference performance remain untouched, the aggregate rack-scale metrics take a noticeable hit, with total memory bandwidth slipping from 1,580 TB/s to 1,400 TB/s.
The eight-GPU HGX Rubin NVL8 server board experiences some of the steepest reductions, losing 26 percent of its memory bandwidth alongside declines in dense training figures. Industry analysts suggest these alterations reflect the realities of manufacturing constraints and supply chain challenges tied to high-bandwidth memory production, prioritizing guaranteed hardware deliveries to major cloud partners over elusive performance ceilings.
These adjustments shift the competitive landscape on paper, particularly as rival manufacturers outline competing hardware specifications. With Nvidia opting for more conservative specifications to maintain hardware rollouts, market observers continue to monitor how these physical adjustments will influence enterprise deployments across data centers.
Comments