NVIDIA Vera Rubin Page Lists 19.2 TB/s HBM4, Not 22
NVIDIA's Vera Rubin NVL72 page lists 19.2 TB/s of HBM4 bandwidth per GPU, below the up to 22 TB/s in its Jan. 5 blog. NVLink per GPU is listed at 3 TB/s, not 3.6.
2 min readAI-assisted
NVIDIA's Vera Rubin NVL72 rack, from the company's product page. The photo shows the rack, not the bandwidth figures.
Image: NVIDIA
Source · Press image from the NVIDIA product page, used for news reporting.
NVIDIA's Vera Rubin NVL72 product page, read on Oct. 26, lists HBM4 bandwidth of 19.2 TB/s per GPU, below the up to 22 TB/s in its Jan. 5 technical blog.
NVIDIA's Vera Rubin page on Oct. 26 lists 19.2 TB/s of HBM4 per GPU, against up to 22 TB/s on Jan. 5. NVLink is 3 TB/s per GPU and 216 TB/s for the rack.
Illustration (AI-generated)
The rack is the one in NVIDIA's DSX POD container, which Voltgrid reported on Oct. 21 from GTC Berlin. NVIDIA said then that one container holds seven Vera Rubin NVL72 racks, 504 GPUs and 252 Vera CPUs, and that a prepared site can have it running in as little as 14 days. That report did not carry the bandwidth figures on the page now.
HBM4 is the memory stacked on each GPU, and NVLink is the link among the 72 GPUs in one rack. Capacity is still listed at 288 GB per GPU, so the published change is how fast that memory and the rack fabric move data, not how much memory is on the chip. Buyers planning factories around the January bandwidth, including that seven-rack container, are now looking at a lower figure on NVIDIA's own table.
What the specification table shows
The table gives the full rack 20.7 TB of HBM4 and 1,400 TB/s of memory bandwidth. The two-GPU superchip is listed at 576 GB and 38.5 TB/s. One GPU stays at 288 GB, with the 19.2 TB/s figure beside it. NVLink on the same rows is 216 TB/s for the rack, 6 TB/s for the superchip and 3 TB/s for a single GPU. Footnote 4 marks networking bandwidth as bi-directional: 32.4 TB/s on the rack, 0.9 TB/s on the superchip and 0.45 TB/s on the GPU.
January's published figures
On Jan. 5, NVIDIA's newsroom said each GPU offers 3.6 TB/s of NVLink bandwidth and that the Vera Rubin NVL72 rack provides 260 TB/s. The technical blog that day said the Rubin GPU carries up to 288 GB of HBM4, with aggregate bandwidth of up to 22 TB/s, and that NVLink 6 delivers 3.6 TB/s of bidirectional bandwidth per GPU. Memory capacity on the current page matches that 288 GB line. The bandwidth lines are lower.
IT Home wrote on Oct. 26 that a June copy of the document had the superchip at 44 TB/s of HBM4 bandwidth and 7.2 TB/s of NVLink, compared with 38.5 TB/s and 6 TB/s now, and that superchip scale-out bandwidth had been 0.8 TB/s. The June numbers are IT Home's reading of an earlier copy. The live-table figures above are taken from NVIDIA's page.
What is still unstated
No NVIDIA newsroom post read for this story explains the lower bandwidth, and the product page gives no reason. TechPowerUp's report the same day tied the change to signal integrity and to posts on X. Those pin-speed and frequency details are not on NVIDIA's page, and this account does not adopt them.
Google Cloud said on Oct. 21 that A5X instances built on Vera Rubin NVL72 are set for general availability in November, as Voltgrid reported. Neither company has said whether that service will ship on the bandwidth numbers now on the product page.
