Another Giant Leap: The Rubin CPX Specialized Accelerator & Rack

NVIDIA's Rubin CPX targets long-context processing: the work of reading a very large prompt before an answer appears. The September 2025 announcement describes a chip with 128GB of GDDR7 memory that works with Vera CPUs and Rubin GPUs on massive-context workloads.

SemiAnalysis explains the design rationale in two stages. Prefill processes the supplied context, while decode generates the response step by step. Prefill can make heavier use of arithmetic capacity, while decode often depends more on memory bandwidth, so cheaper GDDR7 memory suits the first stage and HBM-equipped Rubin GPUs suit the second. The article's cost advantages are estimates rather than measured results from a deployed CPX system.