What changed
Arm has introduced a new methodology for evaluating CPU performance specifically for agentic AI workloads, developed in collaboration with Fujitsu, Meta, and OpenAI. The framework focuses on representative agentic use cases, repeatable CPU measurements, and validation against end-to-end outcomes. This approach aims to better reflect the real-world demands of AI agents, which maintain state, invoke tools, and execute tasks continuously.
What this means in practice
The new benchmarks measure how well CPUs handle the sustained, concurrent workloads of agentic AI, emphasizing throughput and responsiveness rather than raw theoretical capacity. Arm’s AGI CPU demonstrated significant performance improvements over AMD EPYC processors in several key areas. For example, it delivered an average of 1.99x modeled performance per rack compared to AMD EPYC 9755 across seven workloads, driven by higher compute density and power efficiency. In agentic workflows and orchestration, Arm AGI CPU showed an average 2.0x improvement, while reasoning and planning saw a 2.2x boost. The CPU also reduced sandbox creation latency by up to 45% compared to AMD EPYC 9965, enabling faster agent execution.
Limitations and availability
The benchmarks were independently verified by Signal65 against AMD EPYC Turin across tooling, orchestration, and reasoning and planning. Arm is contributing an AGI CPU reference server to the OCP Global Summit, while Meta’s Yosemite YV5 integrates AGI CPU into an OCP rack-scale design. The methodology is designed to evolve as agent architectures and software change, ensuring it remains aligned with industry needs. The full methodology white paper is available for further details.
