Another Giant Leap: The Rubin CPX Specialized Accelerator & Rack

Nvidia has announced the Rubin CPX, a specialized GPU solution designed to optimize the prefill phase of AI inference. This new accelerator emphasizes high compute FLOPS while utilizing less expensive GDDR7 memory for lower memory bandwidth, contrasting with HBM-equipped GPUs optimized for the decode phase. This hardware specialization enables a more efficient disaggregated serving architecture, where different stages of inference are handled by purpose-built chips. The Rubin CPX will be offered in several new rack-scale configurations, including mixed systems with R200 GPUs and dedicated CPX racks. This move is positioned as a significant advancement that increases the architectural gap between Nvidia and its competitors like AMD and custom silicon providers.