Arm Calls for Rack-Level CPU Benchmarks for Agentic AI, Claims Twice AMD's Performance

British semiconductor intellectual property (IP) company Arm said on Oct. 15 that it has developed a new central processing unit (CPU) benchmarking framework for agentic artificial intelligence (AI), with support from Fujitsu, Meta and OpenAI.

The company said the framework is designed to evaluate real-world agentic AI workloads and performance across the entire computing process.

Arm claimed its AGI CPU delivered approximately twice the task throughput per rack of AMD's fifth-generation EPYC Turin 9755 server CPU under the new framework. The measurements were conducted by analyst firm Signal65 at Arm's request. The AGI CPU, unveiled in March, is Arm's first internally designed server CPU and was jointly developed with Meta.

Unlike chatbots that respond to a question and stop, AI agents repeatedly invoke external tools, retrieve data and execute code until a task is completed. GPUs handle model inference, while CPUs manage tool calls, data movement and task scheduling. As the number of agents increases, CPU workloads also grow. Arm said conventional CPU benchmarks and model inference metrics fail to capture these performance demands.

The new framework evaluates tooling, including calls to external applications, databases and services and data processing; reasoning and planning for breaking down complex tasks and organizing execution; the time required to create isolated code execution environments, or sandboxes; and the ability to complete real-world coding tasks and operate in command-line interface (CLI) environments.

Meta is also adopting Arm's AGI CPU in its next-generation open server rack design, Yosemite YV5.