Arm AGI CPU Achieves Nearly 2X Rack Performance Over AMD EPYC 9755 in Agentic AI Workloads

Arm AGI CPU Outperforms AMD in New Agentic AI Benchmarks

Arm announced its AGI CPU delivers nearly twice the rack-level performance of AMD EPYC 9755 across seven agentic AI workloads, according to an independent analysis by Signal651

1

. The results stem from a new CPU evaluation methodology developed by Arm with support from Fujitsu, Meta, and OpenAI, specifically designed to assess CPU performance in multi-step workflows where AI agents execute tools, retrieve data, and maintain state within isolated environments1

1

. The assessment examined workloads involving software tools, autonomous task execution, information retrieval, and support for AI model serving, including Terminal-Bench 2 and SWE-bench2

2

.

Understanding the New CPU Evaluation Methodology

Agentic AI is transforming how CPUs function in modern AI data centers by requiring them to handle complex, multi-step workflows beyond simple inference tasks1

1

. While accelerators handle inference, CPUs execute tools, serve database or API requests, prepare data, and coordinate work for accelerators. Arm argues that CPU performance should be measured by how much useful work a system can sustain while staying responsive, not just by concurrent agent capacity1

1

. The methodology centers on representative agent workloads, repeatable CPU measurements, and validation against end-to-end task outcomes. To address variability in agent behavior, Arm proposes replaying recorded agent actions to separate CPU performance from factors like model inference delays and remote service latency1

1

.

Rack Density Drives Arm's Performance Advantage

The modelled rack-level performance advantage stems significantly from differences in server power consumption, allowing more Arm-based servers within the same electricity allocation2

2

. A 24 kW rack could accommodate 21 to 23 AGI CPU servers compared with 13 dual-socket EPYC 9755 systems, providing between 5,712 and 5,888 physical cores for Arm versus 3,328 for AMD2

2

. Individual server power ranges from 950 to 1,050 watts for Arm's systems, while AMD's dual-socket configurations carry a rated consumption of 1,700 watts2

2

. Across the seven tests, Arm's processor achieved an average performance ratio of 1.99x, with individual results ranging from 1.43x to 3.19x compared with AMD EPYC 97552

2

.

Detailed Test Results and Configuration

For the tests, Arm used two AGI CPU configurations: a 128-core chip at 2.8 GHz with DDR5-7200 for Terminal-Bench 2 and SWE-bench replay, and a 136-core chip at 3.2 GHz with DDR5-8000 for the other five workloads1

1

. The AMD comparison used one socket of the 128-core EPYC 9755 with DDR5-64001

1

. At the chip level, Arm's AGI CPU won five of seven tests, leading by 9% on Terminal-Bench, 13% on SWE-bench, and 97% on fastText, but lagging by 12% and 5% in MySQL and Milvus throughput respectively1

1

. On average, the AGI CPU demonstrated a 20% per-chip advantage in CPU performance1

1

. Across workload categories, Arm recorded modelled rack-level advantages averaging 1.7x for tooling, 2.0x for agentic workflows, and 2.2x for reasoning and planning2

2

.

Sandbox Creation Speed Matters for Tool Execution

Separate testing using Sandbench compared Arm's processor against AMD's EPYC 9965, measuring the time required to establish isolated execution environments for AI agents2

2

. Arm completed cold starts 45% faster, warm starts 27% faster, and its worst-case scenario under fourfold oversubscription 41% faster2

2

. These improvements matter because AI agents frequently create execution environments while accessing services, running code, and processing information, and faster setup can reduce delays in state management and tool execution2

2

.

What This Means for AI Deployments and x86 Infrastructure

Arm executive vice president for Cloud AI Mohamed Awad stated that "The industry needs to measure how well systems handle the work agents actually complete," describing the need for workload-specific performance evaluation2

2

. For AMD and Intel, the findings suggest that Arm-based servers could challenge x86 infrastructure in AI deployments where power limits constrain computing capacity2

2

. Higher rack-level performance could let operators accommodate more agentic AI workloads within existing power allocations, with the report estimating that database and vector-search operations could require 30% to 35% fewer racks for equivalent processing workloads2

2

. However, the results apply to selected processors, workloads, and configuration, and do not establish the AGI CPU as the overall leader for agentic AI1

1

. The findings rely partly on power-based modelling, while testing on final production hardware remains outstanding2

2

. Software compatibility, deployment costs, and real-world operating conditions will influence whether the modelled advantage translates into wider adoption in AI hardware benchmarking2

2

.