NVIDIA Vera CPU Sets New Benchmark for Agentic AI Throughput and Speed

AI agents now perform multi‑step workflows that constantly switch between GPU reasoning and CPU execution. As enterprises deploy fleets of agents, the underlying CPU must deliver both raw speed for individual tasks and the density to host thousands of concurrent agents within a fixed power envelope.

NVIDIA's Vera CPU addresses this dual requirement. In a 100‑megawatt data‑center model, a single Vera chip outperforms the fastest x86 processor on agent latency while supporting 1.6× more agents per watt. Compared with the most energy‑efficient x86 part, Vera doubles agent speed at comparable efficiency.

Early adopters confirm the gains in production‑style workloads: Perplexity saw a 1.5× speed‑up on agentic coding, DeepInfra recorded up to 2.2× faster orchestration, ClickHouse scaled near‑linearly across 176 cores to top the ClickBench leaderboard, and Daytona launched 1,000 agent sandboxes in 10.8 seconds versus 88.4 seconds on x86.

Why It Matters for GPU and AI Infrastructure

GPU‑centric clouds such as AiGpu rely on CPUs that can keep pace with massive parallel inference without becoming a bottleneck. Vera’s high per‑core performance and superior watt‑level density let operators provision fewer sockets, reduce cooling overhead, and allocate more power budget to GPUs — directly improving the cost‑per‑token of large‑scale agent deployments.

With Vera now shipping to partners, the industry has a single‑socket solution that eliminates the traditional trade‑off between speed and density, simplifying hardware procurement for next‑generation AI factories.