Arm's rack benchmark puts CPU density alongside processor speed
Arm's average per-chip advantage of about 20% becomes a modelled 1.99-fold advantage per rack in its agentic-computing comparison with AMD EPYC 9755. The difference comes from the number of complete servers the study puts inside a 24kW power budget, as well as their throughput.
The seven workloads cover activities including tool execution, workflow orchestration, and reasoning and planning. Arm developed the evaluation methodology with support from Fujitsu, Meta and OpenAI.
Signal65's supporting study was commissioned by Arm. It uses measured per-chip results and Arm's rated server-power figures to model rack capacity. Signal65 describes the systems as early AGI CPU hardware, with broader testing on final production systems planned.
The report models a 24kW rack containing 23 or 21 servers for two AGI CPU configurations, compared with 13 two-socket EPYC 9755 servers. The corresponding physical-core counts are 5,888, 5,712 and 3,328. The EPYC configuration also supports simultaneous multithreading, so physical cores and hardware threads are different quantities.
Lower rated server power and higher modelled density amplify the measured throughput at rack level. The near-doubling is therefore not a claim that an individual Arm core executes every task twice as fast.
Why the rack result grows
A modest throughput advantage can support a substantially larger total workload when more complete servers fit within the same electrical envelope. Here, density contributes more to the rack comparison than the average per-chip gain. That conclusion depends on the server-power figures used in the model: memory, motherboard components and other system loads consume part of the rack budget alongside the processor.
Agentic applications repeatedly create execution environments, retrieve information and call software tools. Some of that work takes place on the host CPU even when an accelerator performs model inference. Delays in those operations can lengthen a complete task without changing the model's token-generation speed.
Arm's methodology separates CPU-side measurements from end-to-end outcomes. That separation can help explain whether a faster result comes from tool execution, orchestration or another part of the system. It also prevents a rack-density calculation from being mistaken for a measured deployment of a fully populated rack.
The result is sensitive to both the workload mix and the number of servers that can operate within the rack budget. Signal65 found an AGI CPU per-chip lead in five of seven tests, not all seven. Changing that mix, the server configuration or the power assumptions changes the rack calculation too.
