NVIDIA Rubin CPX

NVIDIA Rubin CPX is a context-phase inference accelerator announced on 9 September 2025 at the AI Infrastructure Summit. It is a distinct class of part from the Rubin datacenter GPU: same Rubin architecture, but a cost-efficient monolithic die paired with 128 GB of GDDR7 instead of HBM4, on the reasoning that the prefill or context phase of long-context inference is compute bound rather than memory-bandwidth bound. NVIDIA published only three hardware figures for it: 30 PFLOPS of NVFP4 compute, 128 GB of GDDR7, and 3x the attention acceleration of GB300 NVL72, alongside hardware video encode and decode. Memory bandwidth, TDP, process node, transistor count, form factor, NVLink and PCIe generation were never disclosed and are stored as NULL here rather than estimated. CPX was designed to sit alongside Rubin GPUs in the Vera Rubin NVL144 CPX rack (144 Rubin CPX, 144 Rubin GPUs, 36 Vera CPUs, 8 exaflops NVFP4, 100 TB of memory, 1.7 PB/s of bandwidth), or in a dedicated CPX compute tray for existing Vera Rubin NVL144 installations. Both the chip and that rack disappeared from NVIDIA's roadmap at GTC 2026.