Nvidia Unveils Rubin CPX (nvidianews.nvidia.com)

NVIDIA has unveiled the Rubin CPX GPU, a groundbreaking processor designed specifically for massive-context AI tasks such as million-token software coding and generative video applications. Integrated within the new Vera Rubin NVL144 CPX platform, the Rubin CPX delivers an astounding 8 exaflops of AI compute, 100TB of fast memory, and 1.7 petabytes per second of memory bandwidth in a single rack. This represents a 7.5x performance boost over previous NVIDIA systems, enabling complex AI models to process ultra-long contexts with unprecedented speed and efficiency—critical for advancing AI-assisted software development and long-format video generation. Technically, Rubin CPX features a monolithic die design with powerful NVFP4 computing resources, 128GB of GDDR7 memory, and 3x faster attention mechanisms than earlier GPUs, allowing models to handle larger context windows without latency loss. It also integrates video decoding and encoding directly on-chip, making it suitable for sophisticated generative video workflows and AI-powered video search. The platform’s scalability is enhanced through compatibility with NVIDIA’s Quantum-X800 InfiniBand fabric and Spectrum-X Ethernet technologies, helping enterprises monetize AI at scale—projected at $5 billion in token revenue per $100 million invested. Industry leaders like Cursor, Runway, and Magic are already exploring Rubin CPX to revolutionize developer productivity, cinematic content creation, and autonomous AI agents. Supported by NVIDIA’s comprehensive AI software stack—including Nemotron multimodal models and the NVIDIA Dynamo platform—the Rubin CPX promises to accelerate next-generation AI applications with enhanced reasoning capabilities and enterprise readiness. Availability is expected by the end of 2026, marking a significant leap forward in AI hardware tailored for long-context understanding and complex generative workloads.