NVIDIA Lowers Rubin GPU HBM and NVLink Bandwidth Specs
NVIDIA's latest documents reveal that the Rubin GPU system's HBM4 memory bandwidth has been reduced from 44TB/s to 38.5TB/s, and NVLink bandwidth from 7.2TB/s to 6TB/s. This adjustment aims to relax hardware requirements to improve HBM4 yield and final production volume, while scale-out network bandwidth sees a slight increase.
Full Report
On October 26, a comparison between NVIDIA's first-party documentation from this October and third-party archives of documents from June reveals that NVIDIA has lowered the HBM and NVLink bandwidth specifications for Rubin GPU systems, while network bandwidth has increased.
For the Vera Rubin superchip, the HBM4 GPU memory bandwidth was previously 44TB/s and is now 38.5TB/s; NVLink bandwidth was previously 7.2TB/s and is now 6TB/s; scale-out network bandwidth was previously 0.8TB/s and is now 0.9TB/s. These changes are largely reflected proportionally in the Vera Rubin NVL4 and Vera Rubin NVL72 configurations, although the scale-out network bandwidth for the Vera Rubin NVL4 has increased from 0.4TB/s to 1TB/s.
Regarding non-superchip Rubin GPU SMX module solutions, the HBM memory for both HGX NVL8 systems has been reduced from 2.3TB at 176TB/s to 2TB at 130TB/s. Compute performance for NVFP4 and FP8/FP6 training has also been downgraded. NVLink Switch bandwidth has decreased from 28.8TB/s to 24TB/s, while bidirectional network bandwidth has increased from 1.6TB/s to 1.8TB/s.
Lowering the bandwidth thresholds may allow more HBM4 memory modules to meet Rubin GPU requirements, thereby improving final yield rates.
Additionally, the memory specifications for all Vera CPUs have been adjusted from "1.5TB LPDDR5X per chip" to "up to 1.5TB LPDDR5X."