NVIDIA has lowered the HBM4 memory bandwidth and NVLink bandwidth listed in its Vera Rubin spec sheets compared with the figures it gave during development: each Rubin GPU goes from 22 to 19.2 TB/s of memory bandwidth and from 3.6 to 3 TB/s of NVLink. The company says, in a statement sent to Wccftech on October 29, 2026, that the earlier numbers were preliminary targets and that the new ones reflect what the system sustains in volume production.
The Vera Rubin spec change in 20 seconds
- The Rubin GPU drops from 22 to 19.2 TB/s of HBM4 and from 3.6 to 3 TB/s of NVLink.
- In the NVL72 rack, memory bandwidth goes from 1,580 to 1,400 TB/s.
- The 288 GB per-GPU capacity is unchanged, and networking bandwidth rises 12.5%.
- NVIDIA says real-world performance is not affected.
The changes appeared a few days ago on product pages and in the Vera Rubin spec sheet, and were picked up by several reports, including one from Wccftech, which compared them with what was presented at GTC 2026. Asked about it, NVIDIA describes the figures published during development as “preliminary architectural targets” and says the final specifications are what the system sustains at qualified production scale.
What numbers have changed
The revision affects every configuration. On each Rubin GPU, HBM4 capacity stays at 288 GB, but memory bandwidth drops 12.7%, from 22 to 19.2 TB/s, and NVLink 6 bandwidth drops 16.7%, from 3.6 to 3 TB/s. On the Vera Rubin superchip, which combines CPU and GPU, memory bandwidth falls from 44 to 38.5 TB/s and NVLink from 7.2 to 6 TB/s, while networking bandwidth rises from 0.8 to 0.9 TB/s.
In the Vera Rubin NVL72 rack, total memory capacity stays at 20.7 TB, with bandwidth going from 1,580 to 1,400 TB/s (−11.4%); NVLink drops from 260 to 216 TB/s and networking rises from 28.8 to 32.4 TB/s. The biggest cut is in the HGX Rubin NVL8, the eight-GPU system: its memory goes from 2.3 to 2 TB and its bandwidth from 176 to 130 TB/s (−26.1%). That configuration also sees lower training throughput, from 280 to 270 petaflops in NVFP4 and from 140 to 130 in FP8/FP6.
NVIDIA’s explanation
NVIDIA argues that the specs of a single component don’t determine application performance, which in its view is decided at the data-center scale through the co-design of models, software, compute, networking, and memory. It says Vera Rubin NVL72 is in full production and already shipping, and stands by its own figures of 30 times more performance per watt and 30 times more tokens per dollar versus GB300 NVL72, measured with production specifications. Those are the company’s own figures.
To back that up, it cites third-party benchmarks already published with the final specifications, such as SemiAnalysis’s AgentX and Artificial Analysis’s AgentPerf. Wccftech adds that the rumor claiming NVIDIA would also cut the Vera CPU’s memory turned out to be false.
Why it matters to infrastructure buyers
Memory bandwidth is one of the factors that weighs most heavily on large-model inference, so the cut will catch the attention of anyone sizing clusters off spec sheets. NVIDIA maintains that the end result doesn’t change, but the gap between what was announced and what ships is a reminder that figures shown for a product still in development are targets, not commitments. It also lands in the middle of fierce competition for AI racks, with AMD already testing its Helios systems at Meta, and as NVIDIA works on approaches like the prototype that brings optics onto the chip interposer.
Frequently asked questions
What changed in NVIDIA Vera Rubin’s specifications?
NVIDIA lowered HBM4 memory bandwidth and NVLink bandwidth: the Rubin GPU goes from 22 to 19.2 TB/s of memory and from 3.6 to 3 TB/s of NVLink. Networking bandwidth rises 12.5%.
Does Vera Rubin’s memory capacity change?
The 288 GB per-GPU and 20.7 TB per NVL72 rack capacities stay the same. It only drops in the HGX Rubin NVL8, from 2.3 to 2 TB.
Why did NVIDIA change the figures?
According to NVIDIA, the earlier numbers were preliminary development targets, and the new ones reflect what the system sustains in volume production.
Does this affect Vera Rubin’s performance?
NVIDIA says no, and stands by its figures versus GB300 NVL72, measured with production specifications. Those are the company’s own figures.
Sources:
