The rivalry between NVIDIA and AMD in the realm of AI chips is reaching new heights. In response to the competitive threat from AMD's Instinct MI455X, NVIDIA is enhancing its next-generation Vera Rubin platform, with a particular focus on boosting memory bandwidth. According to insights from SemiAnalysis, at the upcoming CES 2026 event, NVIDIA plans to unveil an upgraded version of the Vera Rubin NVL72, boasting memory bandwidth specifications that exceed those of the AMD MI455X chip by a notable 10%. The Vera Rubin GPU comes equipped with eight HBM4 memory stacks, collectively delivering an impressive total memory bandwidth of up to 22TB/sec. This represents a significant leap, being 2.75 times greater than the 8TB/sec offered by the previous-generation Blackwell GPU. In comparison, the AMD MI455X is outfitted with 432GB of HBM4 memory, achieving a memory bandwidth of 19.6TB/s. Furthermore, the Vera Rubin platform incorporates six innovative chips, encompassing the Vera CPU and Rubin GPU. This ultra-cooperative design architecture leads to a remarkable enhancement in performance, with inference capabilities being 5 times greater and training performance 3.5 times higher than those of the Blackwell platform.