NVIDIA Vera targets agent loops with 1.8x sustained per-core x86 performance

Agent speed is not only a GPU problem. After a model decides what to do, the CPU often handles the tool call, code execution, data processing, KV-cache-adjacent work, and result checking before the next model step can begin. NVIDIA’s Vera CPU argument is that this in-between work has become a first-order constraint for AI factories.

The design claim is focused: agentic systems need strong single-thread performance under full load, not only more cores. A single agent loop is sequential because each action depends on the previous result. More cores increase parallel throughput, but they do not shorten the next step inside one agent’s chain. NVIDIA says conventional data center CPUs have chased rentable core count and cost per core, often at the expense of per-core speed and memory access.

Vera is built around Olympus, NVIDIA’s custom CPU core. The company says Olympus delivers 50% higher instructions per cycle than Grace. Vera also pairs the cores with up to 1.2TB/s of LPDDR5X memory bandwidth at less than 40 watts of memory power, plus a monolithic compute die with 3.4TB/s of core-to-core bandwidth. NVIDIA says all 88 cores can use the full memory performance of the CPU, and that Vera reaches 1.8x the sustained per-core performance of x86 on loaded agentic execution workloads.

The partner numbers show why infrastructure teams will pay attention. Perplexity tested Vera on a real coding workflow that clones a repository and runs its test suite in sandboxes; NVIDIA says Vera completed the job about 1.5x faster than x86 and started concurrent sandboxes up to 1.9x faster. Starburst measured 3x faster large-scale SQL analytics, while Redpanda measured up to 6x lower latency for real-time streaming. If agent products are measured by completed work, CPU-side waiting becomes lost GPU revenue. Vera reframes the CPU as part of the agent product path, not a background server component.