Nvidia has updated the specifications for its upcoming Rubin GPU systems, according to a comparison between company documentation from June and October. The revisions show a reduction in HBM and NVLink bandwidth, while network bandwidth has seen an increase across several configurations.
For the Vera Rubin superchip, the HBM4 GPU memory bandwidth has been adjusted from 44TB/s to 38.5TB/s, and the NVLink bandwidth has decreased from 7.2TB/s to 6TB/s. Conversely, the scale-out network bandwidth for the superchip has risen from 0.8TB/s to 0.9TB/s.
The Vera Rubin NVL4 system specifically received a significant boost in scale-out network bandwidth, which increased from 0.4TB/s to 1TB/s. These adjustments are largely reflected across the Vera Rubin NVL4 and NVL72 product lines.
In the Rubin GPU SMX module segment, the HGX NVL8 systems saw HBM memory capacity and bandwidth drop from 2.3TB at 176TB/s to 2TB at 130TB/s. Additionally, the NVLink Switch bandwidth was reduced from 28.8TB/s to 24TB/s, while the bidirectional network bandwidth increased from 1.6TB/s to 1.8TB/s.
Training performance metrics for FP4, FP8, and FP6 formats were also lowered for the SMX module solutions. Furthermore, the memory specification for all Vera CPUs has been updated, changing from a fixed 1.5TB of LPDDR5X per unit to a maximum capacity of 1.5TB LPDDR5X.
