Nvidia trims Vera Rubin memory and NVLink bandwidth specs without announcement

Here is what physically changed: nothing visible, only the spec sheet. Each Rubin GPU is now rated at 19.2 TB/s of HBM4 memory bandwidth, down from 22 TB/s, and 3 TB/s of NVLink 6, the chip-to-chip link, down from 3.6 TB/s. That is roughly 13% and 17% less. Capacity holds at 288 GB, and the 50 PFLOPS sparse NVFP4 inference figure is unchanged. ¹

At rack scale, the Vera Rubin NVL72 drops from 1,580 TB/s to 1,400 TB/s of aggregate memory bandwidth and from 260 TB/s to 216 TB/s of NVLink. The eight-GPU HGX Rubin NVL8 board takes the worst hit: 2 TB of HBM4 at 130 TB/s, versus 2.3 TB at 176 TB/s before, a 26% bandwidth cut. ¹

The edit landed between September 1 and September 11, and no announcement followed. ¹ Some pages still show the old numbers, so the math no longer adds up across Nvidia’s own documents. ¹

The likely reason is memory. The HBM4 standard tops out at 8 Gb/s per pin, or 2 TB/s per stack, while Nvidia had been pushing suppliers toward 10 Gb/s and beyond. Reports said SK hynix and Samsung struggled to hit the 22 TB/s target. ¹ Nvidia has not said whether it used smaller stacks or disabled one. ¹

The competitive picture shifts on paper. AMD’s Instinct MI455X lists 23.3 TB/s using 12 stacks at standard speeds, so it never needed exotic memory. Against the revised numbers, AMD’s Helios rack leads by about 19% on memory bandwidth and about a fifth on scale-up bandwidth. None of these figures are benchmarked. ¹

Rubin is still quick: 19.2 TB/s is about 2.4 times Blackwell’s 8 TB/s, short of the 2.8x Nvidia’s blog still claims. ¹

Our read: This looks like a company choosing shipping hardware over a chase for a memory target that suppliers could not deliver. That is sensible engineering, but quiet edits are a poor way to tell customers, and the “10x tokens per megawatt” marketing now rests on figures Nvidia has not said it rerun. ¹

On the bench: if you are budgeting inference racks, ask for measured memory-bound throughput, not the old spec sheet, because bandwidth is the line that decides decode speed.

Footnotes

1. TechRadar, “Nvidia forced to dial back key bandwidth targets on Vera Rubin hardware — tweaks will affect target specifications for memory and interconnect fabric”. Retrieved November 1, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹