Arm AGI CPU benchmarks point to nearly 2x rack-level gains over AMD EPYC Turin in agentic AI workloads
Arm is pushing deeper into the fast-growing agentic AI server market, and its latest AGI CPU benchmarks suggest the company wants to compete directly with high-end x86 data center processors. In new results shown at OCP 2026, Arm compared its AGI CPU against AMD’s EPYC Turin platform, highlighting major advantages in performance per rack, efficiency, and workload scalability.
The results focus on agentic AI workloads, a category that is becoming increasingly important as enterprises deploy AI agents capable of reasoning, planning, using tools, writing code, and operating in sandboxed environments. These workloads are not only GPU-dependent. They also require strong CPU performance, fast memory access, high core efficiency, and predictable scaling across servers and racks.
Arm’s AGI CPU was introduced as the company’s first server-grade CPU silicon designed specifically around agentic AI requirements. It uses third-generation Arm Neoverse cores and has been tuned for high throughput, reduced memory latency, and efficient large-scale deployment. Arm says the processor is built to deliver a balance of per-core speed, rack-scale density, and lower power consumption.
In the latest benchmark set, the Arm AGI CPU was tested across workloads including TerminalBench and SwayBench, with the comparison centered on AMD’s EPYC 9755 Turin processor based on the Zen 5 architecture.
At the chip level, Arm’s AGI CPU delivered an average of 1.20x the performance of the EPYC 9755 across seven tested workloads. AMD still led in some areas, with EPYC ahead in MySQL and Milvus at the socket level. However, Arm’s strongest claim comes at the rack level, where power efficiency and server density significantly change the comparison.
According to the modeled results, a 24 kW rack can fit 21 to 23 Arm AGI CPU servers, compared with 13 dual-socket AMD EPYC 9755 servers. That translates to roughly 5,712 to 5,888 physical Arm cores versus 3,328 physical EPYC cores in the same power envelope.
The reason is power draw. Arm’s model rates an AGI CPU server at around 950 W to 1,050 W, while a dual-socket EPYC 9755 server is rated at about 1,700 W. When the per-chip benchmark advantage is combined with the higher rack density, Arm claims an average 1.99x rack-level performance advantage across the seven workloads.
The individual rack-level results ranged from 1.43x to 3.19x, depending on the workload. Arm also claims advantages across three major agentic AI categories, with modeled rack-level gains of around 1.7x in tooling workloads, 2.0x in end-to-end agentic workflows, and 2.2x in reasoning and planning tasks.
A key part of Arm’s argument is that agentic AI infrastructure is not just about raw CPU speed. For large AI deployments, power and density can be just as important. If a rack can host more servers within the same power limit, total work completed per rack can rise even if the per-socket advantage is smaller.
There is an important memory-related caveat. The benchmark model reportedly sizes memory at 4 GB per hardware thread. Since AMD EPYC supports simultaneous multithreading, that gives the EPYC rack more memory per server, around 2 TB, compared with roughly 1 TB for the Arm AGI CPU server. If memory were instead sized per physical core, the average rack-level advantage would fall from about 1.99x to roughly 1.81x. That still favors Arm in the modeled scenario, but it shows how configuration choices can affect the final result.
Arm also highlighted sandbox creation latency, which is increasingly relevant for AI agents that need isolated environments to run code, test commands, or interact with tools safely. Against a dual-socket AMD EPYC 9965 system, the AGI CPU created sandboxes faster across runc, Firecracker, and gVisor. The reported reductions were 45% lower cold-start time, 27% lower warm-start time, and 41% lower worst-case latency under 4x oversubscription.
This matters because agentic AI systems often spin up many short-lived execution environments. Lower sandbox latency can reduce overhead and improve responsiveness, especially when many agents are operating at the same time.
The results also reinforce a broader shift in server CPU design. While core counts remain important, agentic AI workloads can benefit heavily from stronger per-core performance, fast memory access, and efficient scaling. AI agents may perform many serial or lightly threaded tasks, such as planning, tool invocation, code execution, and environment setup. In those cases, faster cores and lower latency can be just as valuable as having more cores.
Arm’s AGI CPU appears to be designed around that balance. The platform can be tuned for different deployment goals, including higher performance per core or higher total performance per rack. That flexibility could appeal to hyperscalers and cloud providers looking to optimize infrastructure for AI agents without dramatically increasing power and cooling demands.
However, one major point remains: AMD’s next-generation EPYC Venice processors based on Zen 6 were not included in the comparison. Venice is expected to bring significant improvements, including stronger single-core performance and major architectural upgrades. Once those chips are widely available, a fresh round of benchmarks will be needed to show how Arm’s AGI CPU stacks up against AMD’s latest server platform.
For now, Arm’s benchmark presentation sends a clear message. The company is not only targeting traditional server workloads but also positioning its AGI CPU as a serious contender for next-generation agentic AI infrastructure. With claims of nearly 2x rack-level performance, lower power draw, better density, and faster sandbox creation, Arm is making the case that efficiency and scalability may define the next phase of data center CPU competition.
