英伟达本周公布了全新 Vera Rubin 晶片系统的效能基准，该系统由 Vera CPU 与 Rubin GPU 组合而成。公司高层在圣克拉拉总部的技术研讨会上向记者展示了该系统在功耗与效能方面的提升，并强调英伟达正从单纯的 GPU 供应商转型为完整 AI 系统供应商，以应对业界对能驱动 AI 代理的 CPU 日益增长的需求。

Vera Rubin NVL72 系统每瓦可处理的 token 数量是前代 Grace Blackwell 的十倍，记忆体频宽也提升近三倍。该系统采用全液冷设计，大幅减少机架间连接线缆数量，实现所谓「无线缆运算」与「热插拔」功能，将机架安装时间从数小时缩短至数分钟。OpenAI 已率先部署了一组 Vera Rubin 机架。

英伟达计划于今年下半年全面量产 Vera Rubin，早期客户包括微软、OpenAI 和甲骨文。该产品发布正值竞争对手 AMD 年度大会前夕，AMD 也推出了对标的 Helios AI 晶片机架。值得注意的是，英伟达此次在 CPU 架构上舍弃了业界常见的小晶片设计，改用单片式晶片，以提升资料传输效率与记忆体频宽表现。

Nvidia unveiled its new Vera Rubin chip system this week, showcasing performance benchmarks for the GPU-CPU combination ahead of AMD's annual event. Company executives highlighted Nvidia's strategic shift from being primarily a GPU supplier to offering complete AI systems, driven by the growing need for CPUs capable of orchestrating agentic AI workloads, data flows, and networking tasks.

The Vera Rubin NVL72 system delivers ten times the tokens per watt compared to the previous Grace Blackwell generation and nearly triples memory bandwidth. Featuring fully liquid-cooled, cable-free compute with hot-swappable racks, installation time drops from hours to minutes. OpenAI has already deployed a Vera Rubin rack, and Nvidia plans full production in the second half of this year with early customers including Microsoft, OpenAI, and Oracle.

Nvidia's Vera Rubin launch comes just before AMD's annual conference, where AMD revealed its competing Helios AI chip rack. Notably, Nvidia adopted a monolithic chip design for the Vera CPU rather than the chiplet architecture favored by AMD and others, arguing that a single integrated circuit avoids the memory bandwidth and data movement penalties imposed by stitching together multiple chiplets.