Nvidia Quietly Slashes GPU Bandwidth

Nvidia has made unannounced adjustments to the published specifications for its Vera Rubin GPUs, reducing key bandwidth figures. The per-GPU HBM4 bandwidth has been lowered from 22 TB/s to 19.2 TB/s, a decrease of about 13%. Similarly, NVLink 6 bandwidth has been cut by roughly 17%, from 3.6 TB/s to 3 TB/s.

The HGX Rubin NVL8, a standard eight-GPU server board, saw the most significant impact, with its memory bandwidth reduced by 26%, dropping from 176 TB/s to 130 TB/s. These changes were made without a press release, surfacing through updated product pages, suggesting the figures reflect hardware currently shipping.

This bandwidth reduction alters the competitive positioning, at least on paper. AMD's Instinct MI455X now boasts higher per-GPU memory bandwidth, with its Helios platform also leading in scale-up bandwidth. Nvidia's engineers have previously stated that kernel efficiency, rather than peak bandwidth, is crucial for performance, yet these revisions introduce a new dynamic.