Nvidia has launched a GPU with 128GB of GDDR7 RAM but yeah, there's no way it will sell one to us to run games (www.techradar.com)

Nvidia has unveiled the Rubin CPX GPU, a powerful new accelerator featuring 128GB of cutting-edge GDDR7 memory, specifically engineered for enterprise AI workloads rather than gaming. Built on the Rubin architecture, this GPU targets long-context AI inference tasks such as software development, research, and high-definition video generation, boasting up to 30 petaFLOPs of NVFP4 compute and integrated hardware attention acceleration that is three times faster than its predecessor, the GB300 NVL72. Its inclusion of multiple NVENC and NVDEC units further enhances video processing capabilities, marking a significant leap in specialized AI compute power. The Rubin CPX fits into Nvidia’s broader strategy of disaggregated inference, where it focuses on the compute-intensive context phase while other Rubin GPUs and Vera CPUs handle generation, optimizing throughput and lowering inference deployment costs. The flagship Vera Rubin NVL144 CPX rack combines 144 Rubin CPX GPUs, additional Rubin GPUs, and Vera CPUs to deliver a staggering 8 exaFLOPs of NVFP4 compute, 100TB of high-speed memory, and 1.7PB/s of memory bandwidth, interconnected via advanced InfiniBand or Ethernet with ConnectX-9 SuperNICs. Expected to ship in late 2026 following TSMC tape-out, Rubin CPX paves the way for future iterations like Rubin Ultra and Feynman, which promise even higher density, faster memory, and networking enhancements, underscoring Nvidia’s deepening commitment to AI infrastructure innovation.