Nvidia Quietly Cuts Vera Rubin Memory and NVLink Specs

Published October 29, 2026

Nvidia's next-generation Vera Rubin AI platform isn't quite as fast on paper as the company originally advertised. Sometime in late summer, Nvidia revised the official specifications for Rubin GPUs and the systems built around them, lowering both HBM4 memory bandwidth and sixth-generation NVLink interconnect bandwidth. There was no press release or announcement. The change was spotted by comparing archived copies of Nvidia's Vera Rubin NVL72 product page with the live version, and it appears to have been made in early September.

What changed

Each Rubin GPU still carries 288 GB of HBM4, but its listed memory bandwidth has dropped from 22 TB/s to 19.2 TB/s, a cut of roughly 13%. NVLink 6 bandwidth per GPU falls from 3.6 TB/s to 3 TB/s, about 17% lower.

Spec	Original	Revised
HBM4 bandwidth per Rubin GPU	22 TB/s	19.2 TB/s
NVLink 6 bandwidth per Rubin GPU	3.6 TB/s	3 TB/s
HBM4 bandwidth per Vera Rubin Superchip	44 TB/s	38.5 TB/s
HBM4 bandwidth per NVL72 rack	1,580 TB/s	1,400 TB/s
NVLink bandwidth per NVL72 rack	260 TB/s	216 TB/s
HGX Rubin NVL8 HBM4 bandwidth	176 TB/s	130 TB/s

The biggest hit lands on the eight-GPU HGX Rubin NVL8 platform, where aggregate memory bandwidth falls about 26% and listed HBM4 capacity also shrinks from 2.3 TB to 2 TB. Not everything went down, though: scale-out networking bandwidth per Superchip rose from 0.8 TB/s to 0.9 TB/s, lifting the NVL72 rack total from 28.8 TB/s to 32.4 TB/s.

Why it matters

Nvidia hasn't explained the revision, but the most common theory points to HBM4 memory. Hitting 22 TB/s required pin speeds well above 10 Gbps, and reporting from TechPowerUp says the effective target has dropped from 11.7 Gbps to 9.6 Gbps per pin. Relaxing that requirement would let more HBM4 parts from suppliers qualify for Rubin, which could help Nvidia ship more units at a time when memory supply is tight. The NVLink cut is reportedly tied to signal integrity, with the interconnect's target frequency lowered as well.

It's worth keeping this in perspective. These are published specifications, not benchmark results, and Rubin's headline compute figure of 50 PFLOPS of sparse NVFP4 inference per GPU hasn't changed. Nvidia's product page still lists the platform as available now and in full production. Real-world impact will depend on the workload, with memory-bound tasks like large-model inference and communication-heavy multi-GPU jobs the most likely to feel it.

The revision also hands AMD a talking point. AMD's Instinct MI455X is listed with 432 GB of HBM4 at 23.3 TB/s per GPU, and its Helios rack with roughly 31 TB of HBM4 and about 1.67 PB/s of bandwidth, both ahead of the revised Rubin numbers on paper.

Ironically, Nvidia had raised Rubin's memory bandwidth target earlier in the cycle, reportedly to stay ahead of AMD. That bump now appears to have been walked back.