Arm says AGI CPU nearly doubles modelled rack performance over AMD

New benchmark methodology developed with Fujitsu, Meta, and OpenAI frames Arm's lead in power-constrained data centres

By LineZotpaper

Published19 October 2026

Read Time2 min

Arm has announced that its AGI data centre CPU delivers nearly twice the performance per rack as AMD's EPYC 9755 in agentic AI workloads, according to results modeled using a new CPU evaluation methodology developed with support from Fujitsu, Meta, and OpenAI. The analysis, commissioned by Arm and independently verified by Signal65, positions the chip as a leader in handling tool execution, retrieval, and state management tasks crucial for AI agents, though Arm acknowledges the method is not yet an industry standard.

Arm published results from its new agentic AI CPU performance evaluation, claiming its AGI CPU outperformed AMD's EPYC 9755 in five of seven benchmark workloads at the chip level, spanning tooling, agentic workflows, and reasoning.

The company developed the methodology with input from Fujitsu, Meta, and OpenAI to better reflect the demands of agentic AI, where CPUs handle tool execution, data retrieval, and state maintenance in parallel with accelerator-driven inference. Arm argues that traditional benchmarks do not capture this workload properly.

Across workloads including MySQL, NGINX, SWE-bench replay, and Terminal-Bench 2, Arm's AGI CPU demonstrated a 1.20x per-chip advantage over the EPYC 9755 on average. When modeled at the rack level, assuming a 24 kW power budget, Arm's advantage grew to 1.99x on average, driven largely by lower per-server power consumption allowing more servers per rack.

Signal65, an independent analyst firm commissioned by Arm, verified the findings by remotely observing tests and rebuilding the calculations, reproducing all seven results within 0.5 percent. However, Signal65 did not run a separate hardware rerun in its own lab and could only independently confirm BIOS and memory settings for one of the three systems tested.

Other chipmakers such as AMD and Intel make performance claims based on different benchmarks, and Arm's proposed methodology is not yet an industry-agreed standard. Arm's whitepaper sets out a framework for selecting and interpreting these measurements, noting that the industry itself must develop an agreed benchmark suite and scoring approach.

Analysis

Why This Matters

Arm is directly challenging AMD and Intel on the emerging agentic AI workload, which demands a different balance of CPU tasks than traditional cloud computing.
A 2x performance-per-rack claim, if credibly replicated, could shift procurement decisions in data centres where power and space are the primary constraints.
The methodology put forward by Arm, if adopted by the industry, could redefine how CPU vendors benchmark their parts for AI.

Background

Agentic AI workflows rely on CPUs for tool execution, database query handling, sandbox creation, and managing the state of thousands of parallel agents. As these workloads grow, gauging CPU performance by useful work sustained per rack, rather than raw core count, becomes critical. Arm has been expanding its presence in the data centre competing against dominant x86 architectures. The company's proposed benchmark methodology is its clearest attempt to define the terms of engagement for this new segment.

Key Perspectives

Arm: Their AGI CPU is more efficient and powerful for agentic workloads, fitting more servers into a standard rack power budget. They argue their new methodology sets the right standard for evaluating these emerging workloads, validated by an independent analyst.

AMD and other x86 vendors: Compete using their own established benchmarks and real-world deployments. Arm's methodology is self-serving, and the comparison relies on modeled power consumption and specific software configurations that may not reflect typical data centre operations.

Signal65 and Industry Analysts: The results were reproduced within the constraints of the audit, but without a full hardware replication in an independent lab, the findings must be treated as a self-commissioned analysis with inherent modelling uncertainties.

What to Watch

Whether major industry consortiums or hyperscalers adopt or challenge Arm's proposed benchmark methodology.
Real-world performance data from customers deploying the AGI CPU and AMD EPYC 9755 in agentic AI pipelines.
A response from AMD or Intel, potentially publishing their own benchmarks or a critique of Arm's methodology.

Sources

Arm says its AGI CPU delivers nearly 2X the performance per rack of AMD EPYC 9755 in new agentic AI CPU tests · Latest from Tom's Hardware

Checked

10 claims in this story checked against 1 source: 5 supported.

Show each claim

SupportedHeadline and opening: Arm says its AGI CPU delivers nearly twice the performance per rack of AMD's EPYC 9755, based on modelled results.Latest from Tom's Hardware: “Arm says its AGI CPU delivers nearly 2X the performance per rack of AMD EPYC 9755 in new agentic AI CPU tests.”
Partly supportedOpening: the AGI CPU is positioned as a leader in tool execution, retrieval and state management for AI agents.Latest from Tom's Hardware: “The results — verified in an Arm-commissioned independent analysis by Signal65 — do not establish the AGI CPU as the overall leader for agentic AI, as they apply to the selected processors, workloads, and configuration.”
Partly supportedOpening: the analysis was independently verified by Signal65.Latest from Tom's Hardware: “Signal65, an independent analyst firm commissioned by Arm, verified the findings by remotely observing tests and rebuilding the calculations, reproducing all seven results within 0.5 percent.”
SupportedThe AGI CPU outperformed the EPYC 9755 in five of seven benchmark workloads at chip level.Latest from Tom's Hardware: “Arm’s AGI CPU won the majority of comparisons when compared to its AMD counterpart in five of the seven tests.”
Partly supportedAcross the four named workloads (MySQL, NGINX, SWE-bench replay, Terminal-Bench 2), the AGI CPU had a 1.20x per-chip advantage on average.Latest from Tom's Hardware: “On average, the AGI CPU led by 20% (1.20x) per-chip advantage, calculated across the seven normalized performance ratios.”
Partly supportedThe rack-level advantage grew to 1.99x on average, driven largely by lower per-server power letting more servers fit per rack.Latest from Tom's Hardware: “Signal65 also calculated what happens when memory is allocated per physical core. In this setup, AMD’s memory and power allowance falls, and Arm’s average modeled rack advantage drops from 1.99x to approximately 1.81x.”
SupportedSignal65 reproduced all seven results within 0.5 percent.Latest from Tom's Hardware: “A separate review team rebuilt the calculations with its own code and reproduced all seven results within 0.5%.”
Partly supportedSignal65 did not run a separate hardware rerun in its own lab and could only independently confirm BIOS and memory settings for one of the three systems tested.Latest from Tom's Hardware: “In addition, Signal65 could independently confirm BIOS and memory settings for only the EPYC 9965 system. These settings were unavailable for independent verification on the Arm and EPYC 9755 systems.”
SupportedArm's methodology is not yet an industry-agreed standard.Latest from Tom's Hardware: “Arm’s proposed methodology is not yet an industry-agreed standard.”
SupportedArm argues traditional benchmarks do not capture this agentic workload properly.Latest from Tom's Hardware: “Arm argues that traditional benchmarks do not capture this workload properly.”

Correction

20 October 2026An earlier version of this story said Arm's AGI CPU outperformed AMD's EPYC 9755 across all seven benchmark workloads. It outperformed it in five of seven workloads at the chip level; the modeled rack figures showed a lead in all seven.