Another Giant Leap: Rubin CPX—Specialized Accelerator & Rack, Prefill GPU, Disaggregated Power, BOM, and TCO Gains

NVIDIA introduced Rubin CPX, a single-die accelerator aimed squarely at the prefill phase of inference. It prioritizes compute FLOPs over memory bandwidth—complementing decode-oriented parts—and unlocks the full promise of disaggregated serving (specialized hardware for prefill vs. decode). In rack-scale systems, this widens NVIDIA’s advantage; rivals emulating 72-accelerator racks now face another step change and will likely need their own prefill silicon...