NVIDIA Previews Vera CPU Benchmarks for AI Agent Workloads
NVIDIA released benchmark data for its upcoming Vera CPU, demonstrating that it can deliver 1.6x more AI agents per watt and double the agent execution speed compared to top x86 CPUs. As enterprise deployment of AI agents surges, data centers face bottlenecks transitioning from pure generation to multi-step tasks that heavily tax CPUs. This shift highlights how specialized processors can alleviate power and speed limits in next-generation cloud infrastructure. Early testing partners like Perplexity, DeepInfra, and Daytona reported substantial performance gains, with Daytona showing the ability to launch 1,000 agent sandboxes in 10.8 seconds using Vera.
## BACKGROUND
Unlike basic generative AI models that mainly require GPUs for text or image inference, AI agents must actively search databases, write and test code, and run external tools, creating heavy, continuous workloads for server CPUs. Traditional x86 CPUs have long powered standard enterprise servers, but emerging agentic workflows require hardware tightly integrated with modern AI software stacks.