Nvidia Rubin CPX Accelerates Inference Performance and Efficiency (developer.nvidia.com)

NVIDIA has unveiled Rubin CPX, a purpose-built GPU designed to dramatically enhance inference performance and efficiency for long-context AI workloads like software development and HD video generation. As AI models grow increasingly complex—requiring multi-step reasoning, persistent memory, and handling millions of tokens—the demands on infrastructure have escalated, particularly during the compute-intensive “context phase” of inference. Rubin CPX, built on the Rubin architecture, delivers breakthrough capabilities with 30 petaFLOPs of NVFP4 compute, 128 GB of GDDR7 memory, and hardware-accelerated attention mechanisms that triple performance compared to previous NVIDIA GPUs. This innovation is a centerpiece of NVIDIA’s SMART framework, which advocates for disaggregated inference architectures that separate the compute-bound context phase from the memory bandwidth-bound generation phase, enabling tailored optimization. Rubin CPX integrates seamlessly into these infrastructures alongside NVIDIA Vera CPUs and Rubin GPUs, forming the Vera Rubin NVL144 CPX rack. This system aggregates 8 exaFLOPs of compute power, 100 TB of high-speed memory, and enormous bandwidth to support sustained million-token context workloads with exceptional throughput and latency. Orchestrated by NVIDIA Dynamo and leveraging cutting-edge interconnects like Quantum-X800 InfiniBand, this platform sets new industry benchmarks in efficiency, scalability, and return on investment—promising up to 50× ROI and unlocking new potentials for AI applications spanning advanced coding assistants to generative video pipelines.