Arm claims AGI CPU offers twice the rack performance of AMD EPYC 9755 in new tests

Arm announced last week that its new AGI CPU delivers nearly twice the performance per rack as the AMD EPYC 9755 processor. This claim is based on modeled results from tests covering seven agentic AI workloads, using a specific evaluation methodology developed with industry partners.

The company states that these results were verified through an independent analysis commissioned by Arm and conducted by Signal65. According to Tom's Hardware, the findings do not establish the AGI CPU as the overall market leader for agentic AI tasks. The comparison applies strictly to the selected processors, specific workloads, and the configuration used in the testing environment.

New Benchmarking Standards

Arm argues that agentic AI is fundamentally changing the role of CPUs in modern data centers. These systems must now handle complex, multi-step workflows that go beyond simple inference tasks, which are typically handled by accelerators. AI agents execute tools, retrieve data, and maintain state within isolated environments, functioning similarly to virtual users.

The CPU manages these critical surrounding workloads, including running tools, serving database or API requests, and coordinating work for accelerators. As these workloads scale rapidly, Arm contends that performance must be measured by how much useful work a system can sustain while remaining responsive. Supporting more agents simultaneously offers limited benefit if their executions throttle or if accelerators sit idle waiting for the CPU.

Industry Context and Limitations

The performance claims stem from modeled rack results aligned with Arm’s new CPU evaluation methodology. This approach was developed with support from Fujitsu, Meta, and OpenAI. Arm says this methodology should serve as the new standard for assessing performance in agentic AI environments. It centers on representative agent workloads, repeatable CPU measurements, and validation against end-to-end task outcomes.

However, other chipmakers like AMD and Intel make performance claims based on different benchmarks. Physical constraints such as power, cooling, and space also limit how much hardware operators can deploy. This extends the performance question to the rack level, making efficiency per rack a critical metric for data center operators evaluating new infrastructure.

What should you do?

Security and technology professionals should monitor how these new benchmarking standards evolve and whether they become industry-wide norms. Evaluate your current CPU performance metrics to see if they align with the workloads your organization actually runs. Review vendor claims critically, noting the specific workloads and configurations used in their testing.

What to do and how to stay safe: Arm

• Review internal CPU utilization metrics to identify bottlenecks in multi-step AI workflows.
• Compare vendor performance claims against your specific workload requirements rather than raw benchmarks.
• Assess power and cooling constraints in your data center before scaling agentic AI infrastructure.
• Monitor for updates on the adoption of Arm's new evaluation methodology by other industry players.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does Arm's AGI CPU outperform AMD's EPYC 9755 in all scenarios?

No, the claims apply only to the selected processors, workloads, and configurations tested using Arm's specific methodology.

Who verified Arm's performance claims?

Signal65 conducted an independent analysis commissioned by Arm to verify the modeled rack results.

What companies helped develop the new evaluation methodology?

Arm developed the methodology with support from Fujitsu, Meta, and OpenAI.