Science & Tech · AI · 1 hr ago
Arm AGI CPU nearly doubles AMD's modelled rack performance in AI tests
A Signal65 assessment found that Arm's AGI CPU nearly doubled AMD's modelled rack-level performance across seven tests for agentic AI, which uses software to carry out tasks with some autonomy.
The comparison was with AMD's EPYC 9755 processor and included workloads such as running tools, retrieving information and serving AI models.
The estimates assumed a 24-kilowatt rack could hold 21 to 23 Arm servers, compared with 13 dual-socket AMD systems.
Arm's average advantage was 1.99 times, with results ranging from 1.43 to 3.19 times across the tests.
The advantage partly reflects how many servers fit within the same power limit, rather than a direct comparison of individual processors alone.
The results are modelled, and final production hardware has not yet been tested, so they do not show that Arm is faster for all computing or across complete data centers.
If the advantage holds in further testing, Arm servers could let operators run more AI workloads within existing power limits, but software compatibility, costs and real-world conditions will affect adoption.
Arm’s AGI CPU achieved nearly twice the modelled rack-level performance of AMD’s EPYC 9755 across seven agentic AI workloads, according to an assessment by Signal65.
The assessment estimated an average performance ratio of 1.99x, with results ranging from 1.43x to 3.19x.
Under the assessment’s assumptions, a 24-kW rack could hold 21 to 23 Arm servers, compared with 13 dual-socket AMD systems.
The findings rely partly on power-based modelling, and testing on final production hardware remains outstanding.
- Who
- Arm and AMD were compared in an assessment by Signal65, conducted with Arm’s cooperation.
- What
- Arm’s AGI CPU recorded nearly twice the modelled rack-level performance of AMD’s EPYC 9755 across seven agentic AI workloads.
- When
- The assessment’s publication date is 2026-10-19.
- Where
- The comparison modelled performance in 24-kW racks. A physical test location is not stated.
- Why
- The assessment examined how systems handle workloads associated with agentic AI and compared performance under a fixed power budget.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Average performance ratio
- 1.99x for Arm across seven workloads
- Strongest individual result
- 3.19x
- Rack power budget
- 24 kW
- Arm servers per rack
- 21 to 23
- AMD systems per rack
- 13 dual-socket EPYC 9755 systems
- Sandbench cold starts
- Arm completed them 45% faster
What people are saying
The industry needs to measure how well systems handle the work agents actually complete,