Search
NVIDIA has fine-tuned the bandwidth specifications of its Rubin GPU system, opting to decrease both the HBM memory bandwidth and NVLink bandwidth, while simultaneously enhancing the network bandwidth. In particular, the HBM4 memory bandwidth and NVLink bandwidth of the Vera Rubin superchip have undergone a reduction, whereas the bandwidth for horizontal scaling networks has been amplified. The HGX NVL8 system has also seen a decrease in its HBM capacity, bandwidth, and NVLink Switch bandwidth. By lowering the bar for HBM4 memory bandwidth, NVIDIA aims to enable a broader range of memory products to align with the Rubin GPU's specifications, ultimately boosting production yields. Moreover, NVIDIA has revamped the memory specifications of the Vera CPU, now supporting up to 1.5TB of LPDDR5X per chip.
-
C114 Communication Network -
Communication Home
