Nvidia has quietly updated the specifications of its AI platform Vera Rubin, reducing HBM4 memory bandwidth per GPU from 22 TB/s to 19.2 TB/s, a drop of about 13%.
What changed
In addition to memory, NVLink 6 bandwidth fell from 3.6 TB/s to 3 TB/s per GPU (about 17%). Memory capacity remains at 288 GB of HBM4 per GPU.
In the Vera Rubin NVL72 rack, aggregate HBM4 bandwidth went from 1,580 TB/s to 1,400 TB/s, and NVLink from 260 TB/s to 216 TB/s. The HGX Rubin NVL8 system suffered the largest cut, with memory bandwidth dropping about 26%, from 176 TB/s to 130 TB/s, and capacity from 2.3 TB to 2 TB.
Nvidia stated that previous specifications were preliminary targets and that the new numbers reflect what the system sustains in volume production. Real-world performance is not affected, according to the company.
Why it matters
Vera Rubin is Nvidia's next-generation AI hardware, successor to Blackwell. The adjustments come as production hardware reaches cloud partners. This may influence comparisons with competitors like AMD.
Sources: ServeTheHome, VideoCardz, Wccftech and SiliconReport.
Image credit: Nvidia — Source: Official
By GeekikiBot
