Arm states that its AGI CPU delivers twice the performance per rack of the AMD EPYC 9755, based on simulations across seven Agentic AI workloads covering tool workflows and sandbox creation.
However, this figure is not yet conclusive evidence that the AGI CPU is the overall leader, because the results depend on the chips, tests, configurations, and evaluation methodology proposed by Arm itself. The comparison also uses different software environments.
Arm Chips Used in the Tests
This article references two Arm AGI CPU configurations, based on hardware constraints. One is a 128-core model clocked at 2.8 GHz with DDR5-7200 memory, used for Terminal-Bench 2 and SWE-bench replay.
The other configuration is a 136-core model clocked at 3.2 GHz with DDR5-8000 memory, used for the other five workloads. The two chips were therefore not used in every test under identical conditions.
When CPUs Have to Do More Work Around the Model
AI data centers do not let accelerators work alone. CPUs must run tools, query databases or APIs, prepare data, coordinate with accelerators, and manage agent state in isolated environments.
Performance measurement should therefore look beyond how many agents can run simultaneously. It must also consider how much useful work can be sustained while maintaining good responsiveness, since power, cooling, and rack space all limit the number of machines that can be installed.
Where the AGI CPU Fits in Arm’s Overall Lineup
The AGI CPU is Arm’s data center CPU designed specifically for Agentic AI workloads. Its primary role is to support tools, database services, and the coordination tasks behind agent operations.
Based on the available source material, there are no details about where the AGI CPU sits in Arm’s product lineup or what other variants may exist. It should therefore be viewed as a specialized data center product rather than compared with general-purpose CPUs currently on the market.
What the Source Still Does Not Provide for Comparison with Previous Models
| Item | Arm AGI CPU | Previous Arm model |
|---|---|---|
| Product positioning | Data center CPU for agentic AI workloads | Not specified |
| Model and specifications | 128 cores at 2.8 GHz or 136 cores at 3.2 GHz, using DDR5-7200 or DDR5-8000 | Model name or specifications not provided |
| Testing | Tested across 7 workloads against the AMD EPYC 9755 | No test data |
The source therefore does not yet provide enough information to conclude how the product has progressed from one generation to the next, such as how much performance has improved or how much power efficiency has increased. The current comparison is between the AGI CPU and the AMD EPYC 9755, not an earlier Arm model.
From Database Tools to Sandbox Creation
The Arm AGI CPU was tested with transactional databases behind agent tools, making the results relevant to systems that must handle requests and retrieve data during operation.
Another area involved running agent tasks in sandboxed containers. The test covered 72 types of agent tasks, with 128 sets running concurrently, and measured the overhead of creating sandboxes.
For software engineering tasks, the test used 259 SWE-bench replay tasks to evaluate execution according to recorded sequences. For vLLM, the test measured CPU work around model processing while excluding GPU inference, focusing instead on data preparation and system coordination.
How the Arm AGI CPU Compares with the AMD EPYC 9755
| Topic | Arm AGI CPU | AMD EPYC 9755 |
|---|---|---|
| Chip used | AGI CPU with 128 cores at 2.8 GHz or 136 cores at 3.2 GHz | 128-core EPYC 9755 |
| Memory | DDR5-7200 or DDR5-8000 | DDR5-6400 |
| Workloads | Terminal-Bench, SWE-bench replay, fastT, vLLM, Sandbench, database workloads, and reasoning/planning | Same test suite |
| Chip-level results | Won the comparison in most cases, leading in 5 of 7 tests; ahead by 9% in Terminal-Bench, 13% in SWE-bench, and 97% in fastT | Used as the baseline for comparison |
However, these results do not confirm that the AGI CPU is the overall leader in agentic AI, because they are limited to the selected chips, workloads, and configurations. The two sides also use different software environments, and the evaluation is based on a methodology developed by Arm itself.
Strengths and Limitations of the Test Results
The strength of the tests is that they measure work CPUs actually perform in agentic AI, such as tool calls, database operations, sandbox creation, and multi-step workflows. They also consider rack-level results, including power consumption, cooling, and installation space.
The limitation is that the AGI CPU uses two configurations based on the available hardware, while both sides use different software environments. The results therefore depend on the selected chips, workloads, and configurations. In addition, the tests do not directly measure model intelligence or GPU inference speed.
Costs Beyond the Chip
The Arm AGI CPU should not be evaluated solely by chip speed. Power consumption, cooling, and rack space must also be considered, because all of these factors limit how much hardware a provider can install.
If the CPU cannot keep up with the work around the agent, the accelerator may have to wait, leaving resources underutilized. Performance per rack therefore provides a more realistic reflection of the work a system can complete under actual constraints.
However, the source provides no selling price for the AGI CPU or numerical figure for total system cost. The test results also depend on the selected chips, workloads, and configurations, as well as the different software environments. They can therefore support only a partial assessment of cost-effectiveness.
How to Interpret These Results Going Forward
These results should be viewed as evidence from the selected chips, workloads, and configurations, not as confirmation that the Arm AGI CPU is the overall leader in agentic AI, even though Signal65, which was commissioned by Arm, reviewed the results.
Arm proposes replaying agent actions to separate CPU performance from variations in execution paths and latency caused by model inference or remote services. However, the industry still needs benchmark suites and scoring methods agreed upon by all parties.