Signal65 models 1.99× rack performance for Arm AGI CPU vs. EPYC 9755

In an assessment dated Oct. 12, 2026, Signal65 reported that Arm AGI CPU delivered an average 1.99× the modeled rack-level performance of AMD 5th Gen EPYC 9755 across seven CPU workloads. The result covers agent execution, tooling, retrieval and model-serving support; it is a rack model based on rated server power, not a measurement of completed production racks. Signal65’s assessment names Terminal-Bench 2 and SWE-bench among the benchmarks in its workload set.

What the 24 kW rack model compares

Signal65’s model fits 21–23 Arm servers or 13 dual-socket EPYC 9755 servers into a 24 kW rack. It assigns each Arm server a rated power of 950–1,050 W and each AMD server 1,700 W. That power budget lets the model place more Arm systems in the rack, which is part of how it estimates rack-level work. The modeled configurations contain 5,712–5,888 physical Arm CPU cores and 3,328 AMD cores.

Across the seven workloads, Signal65 reports individual modeled rack-performance ratios from 1.43× to 3.19×. Its category averages are 1.7× for tooling, 2.0× for end-to-end agentic workflows, and 2.2× for reasoning and planning. Those results describe this selected workload set and the stated rack assumptions, not a general-purpose ranking of Arm and AMD processors.

The assessment also translates some results into rack-count estimates. For web serving, agent replay and model-serving overhead, Signal65 estimates that about 0.50–0.53 racks could handle the same CPU work as one EPYC 9755 rack. For database and vector-search workloads, it estimates 30%–35% fewer racks for equivalent processing. These are modeled workload-specific estimates, not measured data-center savings.

Sandbench compares a different AMD processor

Signal65’s separate Sandbench test compares Arm with a two-socket AMD EPYC 9965, not the EPYC 9755 used in the rack model. Across sandbox runtimes runc, Firecracker and gVisor, Signal65 reports 45% less cold-start time and 27% less warm-start time for Arm. In a worst-case test with 4× oversubscription, it reports 41% less time.

These figures concern sandbox-start time. They are distinct from the rack-performance ratios and do not describe equivalent gains across all applications.

Arm commissioned the assessment

Signal65 says it worked with Arm and identifies Arm as the research commissioner. TechRadar separately reported that Signal65 independently verified the processor results. The commissioning disclosure and the reported verification are separate facts; neither changes what the rack model measures. Commenters in a Reddit discussion questioned Arm’s involvement in the methodology and the relevance of a new metric.

Signal65 said it planned follow-up testing on final production hardware and additional workloads.