Arm sets out an agentic CPU method and AGI CPU benchmark results
Signal65 used server-level measurements to model rack-level results across seven workloads; Arm reported a 1.99x modeled average for AGI CPU.
Event / data period: 2026-10-13
Published: · Updated:
Arm introduced an agent-focused CPU evaluation method on October 13, with support from Fujitsu, Meta and OpenAI. The company said the same benchmark scripts ran on Arm AGI CPU and AMD EPYC servers, and that Signal65 independently verified the server-level measurements. The framework is intended to combine representative agent use cases, repeatable CPU measurements and checks against end-to-end outcomes.
The rack comparison Arm disclosed is a model derived from measured server performance, rather than a direct measurement of a complete rack. In that model, AGI CPU averaged 1.99 times the rack-level performance of EPYC 9755 across seven workloads; category averages were 1.7 for tooling, 2.0 for agent workflows and orchestration, and 2.2 for reasoning and planning. Arm separately reported up to 45% lower sandbox-creation latency against EPYC 9965, a different comparison system. These findings are limited to the workloads, server measurements and rack assumptions described by Arm.
Data
Workload count in Arm’s benchmark
Number of workloads compared in the rack-level modeling.
| Arm rack-model workloads | 7 |
Methodology
Arm’s announcement reports seven workloads.
Updated:
Sandbox latency result reported by Arm
Upper-bound latency reduction in Arm’s announcement.
| Sandbox creation latency reduction versus EPYC 9965 (upper bound) | 45 |
Methodology
45% is an upper bound against EPYC 9965; EPYC 9755 is the separate rack-performance comparator.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlynewsroom.arm.com ↗