Key points
- Arm states its AGI CPU provides nearly twice the modeled rack performance of AMD EPYC 9755.
- The evaluation methodology was developed with input from Fujitsu, Meta, and OpenAI.
- Signal65 conducted an independent analysis to verify the reported test results.
Arm has announced that its new AGI CPU achieves nearly double the performance per rack compared to AMD’s EPYC 9755 processor. This claim relies on modeled results from seven specific agentic AI workloads. The company developed a new evaluation methodology for these tests with support from Fujitsu, Meta, and OpenAI. Signal65 performed an independent analysis to verify the findings.
By the numbers
- Nearly 2X: The ratio of Arm AGI CPU rack performance to AMD EPYC 9755 rack performance.
- 7: The number of agentic AI workloads used in the performance comparison tests.
- 1: The number of independent firms, Signal65, that verified the analysis results.
What the figures mean
Agentic AI changes how CPUs function in data centers by handling complex, multi-step workflows. These processors manage tasks like running tools, retrieving data, and maintaining state in isolated environments. Arm argues that performance should measure useful work sustained while staying responsive, rather than just concurrent agent counts. If executions throttle or accelerators idle, supporting more agents offers limited benefit. Power, cooling, and space constraints make rack-level efficiency critical for hardware operators.
What should you do?
Managed service providers should verify if current infrastructure contracts account for agentic AI workload characteristics. Teams usually need to assess whether existing CPU capacity can handle increased coordination tasks for AI agents. Check whether vendor performance claims align with your specific workload profiles and power budgets. Ensure that procurement teams understand the difference between raw inference speed and overall task completion rates.
Operational impact
Organizations must evaluate how shifting CPU demands affect data center density and cooling requirements. Procurement teams should question whether vendor benchmarks reflect actual operational workloads rather than theoretical peaks. Operations leaders need to determine if current rack configurations support the sustained throughput required for agentic tasks. IT managers must communicate with facilities teams to ensure power distribution can handle denser, higher-performance server deployments without exceeding thermal limits.
What to do and how to stay safe: Arm
- Review current server procurement specifications to ensure they account for agentic AI coordination overhead.
- Validate vendor performance claims against your specific workload types rather than relying on generic benchmarks.
- Assess data center power and cooling capacity to handle potential increases in rack density and efficiency.
- Monitor industry developments for standardized evaluation methodologies that align with your operational needs.
Step-by-step guide: Shadow AI Explained: 8 Questions Answered for Secure Operations
General security guidance from the BitShield News newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which company verified Arm's performance claims?
Signal65 conducted an independent analysis to verify the results of Arm's AGI CPU performance tests.
Who helped develop the new evaluation methodology?
Arm developed the evaluation methodology with support from Fujitsu, Meta, and OpenAI.
Does this make Arm the overall leader in AI CPUs?
No, the results apply only to selected processors, workloads, and configurations using Arm's specific methodology.
