Nvidia Vera CPU: "Max Single-Threaded CPU at Scale" for AI

Nvidia is highlighting the single-threaded performance of its Vera CPU, positioning it as a key advantage for agentic AI. The company calls Vera a "max single-threaded CPU at scale." Nvidia also revealed its next-generation Rigel Arm CPU cores. These cores will be part of the Rosa CPU and aim to deliver even higher per-core performance than Vera's Olympus core. This improvement comes from better instruction delivery, more L2 cache, and enhanced memory handling. Vera's design features 88 cores with SMT support for 176 total threads. It uses LPDDR5X RAM at 1.2 terabytes per second and boasts 3.4 terabytes per second of core-to-core bandwidth, which Nvidia claims is three times that of other data center CPUs. Nvidia states that AI inference workloads are often limited by single-thread speed. This is because tasks, like a reasoning AI, require sequential steps where each step depends on the previous one. This makes single-thread performance crucial. Ultimately, Nvidia believes this focus on single-threaded performance in Vera provides a significant edge for demanding AI applications.