NVIDIA introduces Vera, a new CPU class optimized for agentic AI workloads with maximum single-threaded performance per

NVIDIA has unveiled Vera, a CPU designed specifically for the agentic AI era. Unlike conventional data center CPUs built to maximize core count while minimizing cost per core, Vera prioritizes single-threaded performance to accelerate agent loops where each step depends sequentially on the previous one.

The problem Vera addresses is that traditional data center CPUs sacrifice per-core performance for lower cost, using chiplet architectures and reduced memory fabrics. In agentic workloads, however, speed matters critically. Agents operate in continuous loops where the CPU executes tool calls, code execution, data processing and result verification between model calls. Since each agent step must complete before the next begins, faster individual cores determine loop speed, not just total core count. GPU utilization in AI factories is constrained by CPU performance, so any CPU slowdown reduces data center revenue.

At Vera's core is Olympus, NVIDIA's custom ARM v9.2 CPU core, delivering 50 percent higher instructions per cycle than NVIDIA's Grace processor. Vera features up to 88 cores paired with up to 1.2 terabytes per second of LPDDR5X memory bandwidth at less than 40 watts of memory power. A monolithic compute die architecture provides 3.4 terabytes per second of core-to-core bandwidth, three times greater than competing data center CPUs, eliminating memory bottlenecks that slow individual cores.

In loaded CPU workloads representing agentic execution, Vera delivers 1.8 times the sustained per-core performance of x86 processors. Perplexity tested Vera on real coding workflows including repository cloning and test suite execution in sandboxes, finding Vera completed jobs about 1.5 times faster than x86 and started concurrent sandboxes up to 1.9 times faster. Perplexity plans to deploy Vera in its upcoming production system. For data workloads, partners measured three times faster large-scale SQL analytics with Starburst and up to six times lower latency on real-time streaming with Redpanda compared to leading x86 server CPUs.

NVIDIA plans to continue this roadmap with Rosa, its next-generation CPU featuring the Rigel core, which will deliver higher per-core performance than Olympus while maintaining the same silicon footprint through improvements in instruction delivery, L2 cache size and memory efficiency.