Arm announced that its AGI CPU delivers nearly double the per-rack performance of AMD's EPYC 9755 in agentic-AI workloads. The claim comes from modeled rack results across seven tests, using a CPU evaluation methodology Arm developed with support from Fujitsu, Meta, and OpenAI. An independent analysis by Signal65, commissioned by Arm, verified the results.
Arm argues that agentic AI shifts CPU performance from raw concurrency to useful work: systems must run tools, serve database and API requests, and coordinate accelerators without stalling. Its methodology measures representative workloads, keeps CPU measurements repeatable by replaying recorded agent actions, and checks whether CPU improvements actually help end-to-end tasks. The tests include MySQL, NGINX, Milvus, Terminal-Bench 2, SWE-bench replay, fastText, and vLLM CPU overhead, with the vLLM test isolating CPU work around model execution rather than GPU inference.
At the chip level, Arm's AGI CPU won five of the seven comparisons. It led by 9% on Terminal-Bench, 13% on SWE-bench, and 97% on fastText, while trailing by 12% in MySQL and 5% in Milvus throughput. Averaged across the seven normalized results, Arm reports a 20% per-chip advantage; modeled at the rack level, that becomes nearly 2x.
The results do not establish the AGI CPU as the overall agentic-AI leader. They apply only to the selected processors, workloads, and configurations, and they use Arm's own methodology rather than the different benchmarks AMD and Intel often cite. Arm proposes its approach as a new standard, but leaving the industry to develop an agreed benchmark suite and scoring scheme.