Michael JohnsonIn an agentic system, the CPU handles work around the model: managing state and queues, calling tools, processing data, and launching runtime environments. On October 13, Arm introduced a methodology for evaluating CPUs on these workloads, with support from Fujitsu, Meta, and OpenAI. Alongside it, the company published a comparison of the Arm AGI CPU with AMD EPYC processors.
For infrastructure operators, what matters is not only the performance of an individual chip but also the amount of work a server rack can handle within specified power and space limits. In Signal65’s analysis, the AGI CPU delivered an average modeled per-rack performance 1.99 times that of the AMD EPYC 9755 across seven workloads. The figure applies to a modeled full rack and the selected tests.
What Arm proposes measuring
The methodology divides agentic-system work into framework functions—such as state management, routing, scheduling, and telemetry—and tool tasks, including compilation, code execution, search, and database operations. Arm’s methodology paper proposes two complementary approaches: component-level measurements for repeatable CPU comparisons and end-to-end tests to assess task completion.
This distinction helps identify where latency arises: in the CPU, model, network, or application. Component tests can isolate individual operations, while end-to-end tests can show the time and throughput of an entire task. Arm presents the paper as a basis for defining measurement boundaries, workload categories, and areas where the industry still needs to agree on a common approach.
How the comparison results were obtained
Signal65’s analysis covers seven CPU workloads, including agentic execution, tools, search, and model-serving support. Signal65 reports an average modeled per-rack performance ratio of 1.99× versus the EPYC 9755; results across the seven workloads ranged from 1.43× to 3.19×. The rack calculation uses measured server performance and nominal system power: its model fits 21–23 Arm servers in a 24 kW rack, compared with 13 AMD servers.
Across three groups of agentic tasks, Signal65 reports average per-rack ratios of 1.7× for tool use, 2.0× for agent workflows and orchestration, and 2.2× for reasoning and planning. In a separate test of isolated-environment creation, the AGI CPU was compared with the AMD EPYC 9965 instead: Signal65 reports a 45% reduction in cold-start time, 27% in warm-start time, and 41% in the described case with fourfold oversubscription. This is a separate comparison against a different processor using different metrics.
What to consider when choosing servers
Even when inference runs on an accelerator, the CPU handles a significant part of the agentic cycle: running code and tools, processing data, and coordinating parallel tasks. Under heavy contention, these operations can affect execution time and accelerator utilization.
Teams choosing servers should profile their own workflows: which tools are called, how often runtime environments are created, how many tasks run concurrently, and what latency is acceptable. They can then measure repeatable CPU operations separately from the full workflow under comparable power limits. Such results will better reflect their specific infrastructure than applying a ratio from someone else’s test.
Signal65 notes that the results were obtained on early AGI CPU systems and calculated at rack scale using server measurements and nominal power; the company plans further tests on production hardware and a broader range of workloads. Arm’s announcement is therefore important primarily as a proposal to measure the CPU contribution to agentic systems through a combination of repeatable components and completed tasks. The 1.99× figure is a reference point for the modeled scenario described, not a universal performance estimate for every rack.