Nvidia forced to dial back key bandwidth targets on Vera Rubin hardware — tweaks will affect target specifications for memory and interconnect fabric
HBM4 bandwidth dropped 13% while NVLink dropped 17%, and it took six weeks for anyone to notice.
Nvidia has reduced the listed HBM4 bandwidth for its Vera Rubin system, cutting it from 22 TB/s to 19.2 TB/s per GPU and NVLink from 3.6 TB/s to 3 TB/s. The reduction affects the eight-GPU HGX Rubin NVL8 the most, causing a 26% loss in memory bandwidth and a drop in HBM4 capacity from 2.3 TB to 2 TB. AMD's MI455X now leads in per-GPU memory bandwidth, while Helios shows superiority in scale-up bandwidth.
Despite these changes, Nvidia's headline performance remains unchanged. The downgrades were implemented quietly, without a press release or explanation, and surfaced around September 2025, just as the fall production shipments for Vera Rubin were due to begin. This suggests the revised figures pertain to the hardware being delivered rather than a future goal.
The change might be due to challenges in meeting the desired memory bandwidth, possibly resulting in smaller HBM4 stacks per GPU or disabling one of the usual eight stacks. The retreat impacts AMD's position, making its MI455X and Helios platforms appear more competitive in terms of memory bandwidth.
Written by urgent.news from TechRadar's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
