Arm AGI CPU Outperforms AMD in Agentic AI Rack-Level Benchmarks
Arm's architecture leverages power efficiency to pack more cores into standard data center racks.
1Performance Benchmark Overview
Arm’s AGI CPU demonstrated nearly double the modelled rack-level performance compared to AMD’s EPYC 9755 across seven specific agentic AI workloads. Conducted by Signal65, the assessment utilized power-based modelling to estimate efficiency within a fixed 24-kW power budget. These results highlight potential shifts in how data centers evaluate hardware for autonomous AI tasks.
2Rack Density Advantage
The primary performance gain stems from superior rack density enabled by lower power consumption per server. A 24-kW rack accommodates 21 to 23 Arm servers, compared to only 13 dual-socket AMD systems. This configuration allows Arm to deploy between 5,712 and 5,888 physical cores, significantly surpassing the 3,328 cores available in the AMD setup.
3Workload-Specific Efficiency Gains
Arm recorded varying performance advantages across different categories, averaging 1.7x for tooling, 2.0x for agentic workflows, and 2.2x for reasoning and planning tasks. Furthermore, the report suggests that database and vector-search operations could potentially require 30% to 35% fewer racks to achieve equivalent processing capacity, offering a clear path toward optimizing data center footprints.
4Execution Environment Speed
Testing with Sandbench revealed that Arm processors complete isolated execution environment setups significantly faster than x86 alternatives. Cold starts were 45% faster, while warm starts improved by 27%. Under fourfold oversubscription, Arm maintained a 41% speed advantage, which is critical for AI agents that frequently create and destroy environments during complex task execution.
5Industry Collaboration Framework
Arm developed this evaluation framework with strategic support from Fujitsu, Meta, and OpenAI to better reflect real-world agentic AI demands. Mohamed Awad, Arm’s executive vice president for Cloud AI, emphasized the necessity of measuring how systems handle actual agent-driven work. Signal65 independently verified the processor results to ensure the integrity of the comparative data.
6Limitations of Current Modelling
The findings rely on modelled performance rather than final production hardware, meaning real-world results may vary. Arm acknowledged that further testing on finalized systems is required to cover additional workloads. The current data does not demonstrate equivalent gains across general-purpose computing, nor does it suggest an immediate need to replace existing x86 server fleets.
7The x86 Competitive Challenge
For incumbents like AMD and Intel, these findings signal a growing threat in agentic AI deployments where power limits constrain total capacity. If Arm’s modelled advantages hold in production, operators could accommodate significantly more AI workloads within existing power allocations. This creates a compelling case for Arm in power-constrained environments, despite current x86 dominance.
8Future Adoption Factors
While the rack-level performance advantage is notable, wider adoption will depend on software compatibility, deployment costs, and real-world operating conditions. The industry must weigh these hardware gains against the complexities of migrating existing infrastructure. Future benchmarks will be essential to determine if these performance metrics translate into sustained competitive advantages for Arm-based data centers.