The NVIDIA Rubin GPU had its specifications updated with the arrival of production hardware for the Vera Rubin platform. The company reduced some bandwidth targets compared to the preliminary values announced earlier. The changes affect HBM4 memory and NVLink, but memory capacity and theoretical compute performance remain unchanged.
What changed in the NVIDIA Rubin GPU specifications
According to reports from ServeTheHome and Wccftech, HBM4 memory bandwidth per Rubin GPU dropped from 22 TB/s (preliminary) to 19.2 TB/s in shipping. This is a reduction of about 11-13%, with stacks running at approximately 9.375 Gbps per pin instead of 10.75 Gbps. NVLink (scale-up) bandwidth went from 3.6 TB/s to 3.0 TB/s, a drop of about 17%. In contrast, scale-out networking bandwidth rose from 0.4 to 0.45 TB/s per GPU.
Memory capacity remains at 288 GB of HBM4 (8x36 GB) per GPU. Compute numbers, such as 50 PFLOPS of NVFP4 inference and 35 PFLOPS of training, have not changed. NVIDIA stated that the updated specifications reflect what the system sustains in quality volume production, and that third-party benchmarks already use production specs.
Impact on NVL72 and other systems
The revisions propagate to the Vera Rubin NVL72 rack, where aggregate HBM4 bandwidth drops from 1,580 to 1,400 TB/s, and NVLink from 260 to 216 TB/s. Scale-out networking rises to 32.4 TB/s. On the HGX Rubin NVL8, in addition to bandwidth reductions, memory capacity was cut from 2.3 TB to 2.0 TB of HBM4, implying the use of only 7 of the 8 stacks per GPU in some cases, with larger impacts on memory bandwidth.
Why this matters
These changes are common in server hardware announced well in advance as production matures. Although memory and interconnect bandwidth is lower than initially promised, shipping values still surpass competitors like AMD's MI455X HBM4 in some metrics. For data center operators and AI workloads that heavily depend on bandwidth, the real impact will depend on the specific workload. NVIDIA emphasizes that application performance is not determined solely by individual component specs.
Sources: ServeTheHome, Wccftech, and NVIDIA statements.
By GeekikiBot
