Arm AGI CPU racks up 23 servers in 24kW benchmark
Signal65 models nearly twice the rack-level work from Arm's first production CPU, while warning that final hardware testing is still to come.
Arm's AGI CPU is being pitched as a way to squeeze more agentic-AI work into the same data-centre power budget. A new Signal65 assessment, conducted with Arm's cooperation, models up to 23 Arm servers in a 24 kW rack, compared with 13 dual-socket AMD EPYC 9755 systems. The result is an average 1.99x rack-level performance advantage across seven workloads — but the comparison is partly modelled, not a final production-system bake-off.1
What the benchmark found
Signal65 tested the Arm AGI CPU against AMD's fifth-generation EPYC 9755 across workloads covering agent execution, tooling, information retrieval and support for AI model serving. The assessment included Terminal-Bench 2 and SWE-bench, alongside other infrastructure tests. Individual rack-level results ranged from 1.43x to 3.19x in Arm's favour, with the strongest gains appearing in reasoning and planning workloads.1
The rack comparison is where the headline result comes from. Signal65's power assumptions put Arm systems at 950 to 1,050 watts each, while the two-socket AMD configuration is rated at 1,700 watts. That allows the modelled 24 kW rack to hold 21 to 23 Arm systems, delivering between 5,712 and 5,888 physical cores, versus 3,328 cores for the AMD setup.1
The same report says Arm's chip delivered 45% faster cold sandbox creation, 27% faster warm starts and a 41% improvement in its worst case under fourfold oversubscription when compared with AMD's EPYC 9965 in Sandbench testing. Those figures matter for agent systems that repeatedly create isolated environments, but they should not be read as a claim that Arm has won every server workload.1
Arm's first production CPU
Arm describes the AGI CPU as its first production silicon, built around Neoverse V3 cores for continuously running agentic workloads. The company lists up to 136 cores, a 300-watt TDP, DDR5-8800 memory support, 96 PCIe Gen6 lanes and CXL 3.0 connectivity on its product page.2
Arm's launch announcement says the chip can deliver more than twice the performance per rack of x86 CPUs and supports high-density 1U systems. It also names Meta as a lead partner, with Cloudflare, OpenAI, SAP, SK Telecom and other companies among the wider ecosystem supporting deployments.3
The important qualification is that Signal65 says its rack results are calculated from measured per-chip results and Arm's rated server power. The analyst plans follow-up testing on final production hardware, with broader competitive comparisons. Power, software compatibility, deployment cost and real-world operating conditions will decide whether the modelled density advantage survives outside the benchmark.1
Our opinion
This is a meaningful infrastructure story because rack power, not just chip speed, is becoming a bottleneck for AI deployments. The Arm AGI CPU's apparent advantage is strongest when the question is how much agentic work fits inside a fixed electricity envelope. But the 24 kW result is a model built from early systems and rated power assumptions, so it is evidence of a promising direction rather than a universal AMD-killer verdict. The follow-up production testing will matter more than the neat 2x headline.