Arm claims its AGI CPU delivers nearly twice AMD’s EPYC Turin performance per rack, thanks to its higher server density.
Research outfit Signal65 tested early Arm systems against AMD’s Zen 5-based EPYC 9755 across seven workloads. Arm commissioned the study, and it seems reasonably pleased with its purchase.
The headline result is an average of 1.99 times the modelled performance per rack, with individual results ranging from 1.43 to 3.19 times AMD’s performance. That distinction between measured processor results and modelled racks deserves attention before anyone orders several thousand servers.
Per chip, Arm averaged 1.20 times the EPYC 9755’s performance. AMD won the MySQL and Milvus tests, where Arm scored 0.88 and 0.95 times its rival, respectively. The larger rack-level advantage comes from fitting more Arm systems within the same electricity budget.
Arm’s rated-power model places 21 to 23 AGI CPU servers in a 24kW rack, compared with 13 dual-socket EPYC 9755 servers. The Arm configurations have between 5,712 and 5,888 physical cores per rack, compared with AMD’s 3,328. Rated server power ranges from 950W to 1,050W for Arm, compared with 1,700W for the AMD systems.
Packing more processors into the available power envelope creates a useful commercial advantage, even if the headline makes it sound like every Arm chip has eaten two EPYCs for breakfast.
Memory allocation affects the comparison. The model provides 4GB per hardware thread, giving the AMD servers approximately 2TB each because they support simultaneous multithreading. Arm’s servers receive approximately 1TB. Allocating memory per physical core would reduce Arm’s average rack-level advantage from 1.99 to about 1.81.
The seven workloads span MySQL, NGINX, Terminal-Bench 2, SWE-bench replay, Milvus vector search, vLLM CPU overhead and fastText. Arm’s modelled rack results average 1.7 times AMD’s performance for tooling, twice its performance for agent workflows and 2.2 times for reasoning and planning support.
A separate Sandbench comparison used a dual-socket EPYC 9965 system. Across runc, Firecracker and gVisor, Arm reported sandbox creation taking 45 per cent less time for cold starts and 27 per cent less for warm starts. Worst-case creation time fell by 41 per cent when workloads were oversubscribed by a factor of four.
The AGI CPU uses Neoverse V3 cores, with Arm emphasising memory bandwidth, responsiveness and sustained performance when many agents compete for resources. Arm has developed its evaluation methodology with support from Fujitsu, Meta and OpenAI. The approach combines repeatable CPU measurements with checks against completed agent tasks.
The comparison excludes AMD’s Zen 6-based EPYC Venice processors. Signal65 plans further testing on final production Arm hardware, with more workloads and competing systems.
