Nvidia revamps rubin cpx chip for enhanced AI inference performance

According to analyst郭明錤, NVIDIA has restarted the Rubin CPX project, focusing on AI inference pre-filling acceleration. The new design shows significant changes, with production expected to begin in the first quarter of 2027. The new Rubin CPX chip offers nearly the same computational power as the standard Rubin GPU, with improved pre-filling performance and a maximum power consumption of 2300W. It is equipped with 168GB of HBM4 memory, an upgrade from the previous 128GB GDDR7. At the rack level, eight new Rubin CPX chips form a compute tray, with eight trays making up a rack module. These modules use Spectrum-6 full copper Ethernet for expansion. The new Rubin CPX uses an independent MGX ETL rack, with each rack containing 1 to 4 rack modules. The rack modules are connected via Spectrum-6 and OSFP optical solutions. NVIDIA recommends pairing Rubin CPX with standard Rubin chips in a 1:1 ratio, and connecting Rubin CPX ETL racks with Vera Rubin NVL72 racks using Ethernet-based RDMA. 郭明錤 noted that over 50% of current AI inference workloads involve processing input and building KV Cache. The new Rubin CPX can handle pre-filling loads with more flexible deployment and lower costs. The compute tray has 1.34TB of HBM memory, sufficient for most long-context pre-filling and KV Cache requirements.