NVIDIA 推出面向推理 prefill 阶段的 Rubin CPX,以更高算力占比和低成本内存减少 HBM 与 NVLink 资源浪费。单芯片提供 20 PFLOPS 稠密 FP4 算力、128GB GDDR7 和 2TB/s 带宽,并通过 PCIe Gen 6 联网。该产品将用于固定配置的 VR200 NVL144 CPX 或更灵活的 Vera Rubin CPX 双机架方案。 · NVIDIA introduced Rubin CPX for the prefill phase of inference, using a compute-heavy design and lower-cost memory to reduce wasted HBM and NVLink resources. Each chip provides 20 PFLOPS of dense FP4 compute, 128GB of GDDR7, and 2TB/s of memory bandwidth, with PCIe Gen 6 connectivity. It will be deploye