Arm announced results from independent tests showing its new AGI CPU delivering nearly double the performance per rack compared to the AMD EPYC 9755 on agentic artificial intelligence workloads.
What happened?
According to a Signal65 analysis commissioned by Arm and developed with support from Fujitsu, Meta and OpenAI, the Arm AGI CPU achieved an average of 1.99x modeled performance per rack across seven specific AI agent workloads. Tests include MySQL, NGINX, vector search with Milvus, Terminal-Bench 2, SWE-bench replay, fastText and vLLM overhead.
The advantage of the Arm AGI CPU comes mainly from higher density: a 24 kW rack fits 21 to 23 AGI CPU servers (5,712 to 5,888 cores) versus 13 dual-socket EPYC 9755 systems (3,328 cores). Arm server power is between 950 and 1,050 W versus 1,700 W for AMD.
Why does it matter?
AI agents run tasks continuously and in parallel, requiring efficient CPUs for orchestration, tools and reasoning, not just GPUs for inference. Arm positions the AGI CPU as the company's first production silicon, based on Neoverse V3, focused on data center scale for the agent era.
The results of the Arm AGI CPU do not claim absolute leadership, as they depend on chosen configurations and workloads, but highlight efficiency in power- and space-constrained environments. Arm reduces sandbox creation latency by up to 45% compared to the EPYC 9965.
Context
The Arm AGI CPU was announced in March 2026 as the foundation for agentic AI infrastructure. Partners such as OpenAI, Meta and others are already evaluating its use. Commercial systems are available via ASRockRack, Lenovo and Supermicro.
Sources: Tom's Hardware, Arm Newsroom.
By GeekikiBot