Arm AGI CPU Nearly Doubles AMD EPYC 9755 in Rack-Level AI Test
Signal65, working with Arm, modelled a 24 kW rack where Arm's AGI CPU held 21 to 23 servers against 13 dual-socket EPYC 9755 systems.
Signal65 has published an assessment, conducted with Arm's cooperation, finding that Arm's AGI CPU delivers close to double the modelled rack-level performance of AMD's EPYC 9755 across seven agentic AI workloads, according to TechRadar. The Arm processor averaged a 1.99x modelled rack-level performance ratio, with individual workloads ranging from 1.43x to 3.19x.
The comparison was built around a 24 kW rack. Within that power envelope, the assessment modelled 21 to 23 Arm AGI CPU servers against 13 dual-socket EPYC 9755 systems. Arm systems were rated at 950 to 1,050 watts each, while AMD's dual-socket configurations were rated at 1,700 watts.
That configuration gives the Arm rack 5,712 to 5,888 physical cores, compared with 3,328 for AMD. Modelled rack-level advantages averaged 1.7x for tooling, 2.0x for agentic workflows and 2.2x for reasoning and planning. The report also estimated that database and vector-search operations could need 30% to 35% fewer racks for equivalent processing.
In separate Sandbench testing against AMD's EPYC 9965, Arm completed cold starts 45% faster and warm starts 27% faster, with its worst case 41% faster. The workloads covered included Terminal-Bench 2 and SWE-bench, spanning software tools, autonomous task execution, information retrieval and AI model serving.
Arm developed its evaluation framework with support from Fujitsu, Meta and OpenAI, and Signal65 independently verified the processor results. Mohamed Awad, Arm executive vice president for Cloud AI, said the industry needs to measure how well systems handle the work agents actually complete.
Testing on final production hardware remains outstanding. Arm said further testing would cover additional workloads and competitors.
Quick answers
What did Signal65 find in its Arm AGI CPU assessment?
It found Arm's AGI CPU averaged a 1.99x modelled rack-level performance ratio versus AMD's EPYC 9755 across seven agentic AI workloads, with individual workloads ranging from 1.43x to 3.19x.
Was the assessment run on production hardware?
No. Testing on final production hardware remains outstanding, and Arm said further testing would cover additional workloads and competitors.
Who supported Arm's evaluation framework?
Arm developed the framework with support from Fujitsu, Meta and OpenAI, and Signal65 independently verified the processor results.