NVIDIA adjusts Vera Rubin specifications to accommodate HBM4 production

NVIDIA has reportedly revised the technical specifications for its upcoming Vera Rubin superchip architecture. According to recent documentation compared to earlier June reports, the company has lowered the HBM4 memory bandwidth from 44TB/s to 38.5TB/s, representing a decrease of approximately 12.5%.

NVLink interconnect bandwidth has also been adjusted, dropping from 7.2TB/s to 6TB/s. For individual GPUs, NVLink bandwidth moved from 1.8TB/s to 1.5TB/s. However, NVIDIA has increased horizontal scaling network bandwidth from 0.8TB/s to 0.9TB/s, which aids communication efficiency across large-scale AI clusters.

Industry analysts suggest these adjustments are intended to lower the technical threshold for HBM4 memory. By doing so, NVIDIA may allow a broader range of products from suppliers like Samsung and SK Hynix to meet the Rubin GPU specifications, thereby securing more stable production volumes as HBM4 supply remains in the ramp-up phase.