AI
Nvidia revives Rubin CPX accelerator with HBM4 and standalone rack design
TF International analyst Ming-Chi Kuo says Nvidia has revived its Rubin CPX inference accelerator with a major redesign, including 168 GB of HBM4 memory per GPU and a move to a standalone MGX ETL rack. Production is now slated to begin in Q1 2027, with CPX handling the prefill stage of inference alongside Vera Rubin NVL72 systems via Ethernet RDMA at a recommended 1:1 ratio.