Nvidia's new Vera Rubin CPU targets agentic AI workloads
nvidia unveiled its vera rubin cpu-gpu platform, claiming 10x better token-per-watt efficiency than grace blackwell. the shift addresses growing demand for cpus to orchestrate complex ai agents.
nvidia revealed new details for its vera rubin platform ahead of amd's product event, emphasizing a system architecture designed to support the shift from passive models to active ai agents. the nvlink72 system integrates 72 gpus with 36 cpus, offering localized memory subsystems with nearly three times the bandwidth of blackwell and a 'hot-swappable' design that reduces cable complexity. benchmarks presented by nvidia executives indicate the system delivers ten times the tokens-per-watt efficiency of grace blackwell and outperforms competitors in cpu-based agentic tasks. openai is already testing the new racks, as nvidia attempts to capture the market for data orchestration hardware currently dominated by traditional cpu suppliers. the company plans to start selling the standalone vera cpu to chinese customers as early as august.