Arm says its AGI CPU averaged nearly twice the performance per rack of AMD’s EPYC 9755 across seven AI-related workloads. The useful detail is that the rack advantage is modelled from power assumptions and server counts, not measured on two fully populated racks.
Arm Watch analysis
What happened
Tom’s Hardware reports that Arm compared its AGI CPU with AMD’s EPYC 9755 on seven workloads, including database and vector-search tests, agent task benchmarks and CPU overhead for vLLM. The article says Signal65 reviewed Arm’s test runs, logs and calculations, and reproduced the seven results within 0.5%.
At chip level, Arm averaged a 1.20× advantage across the tests, winning five of seven. The reported 1.99× average per-rack result comes from modelling how many servers fit within a 24 kW rack budget, using estimated server power. Arm’s estimate allows 23 servers for one configuration, against 13 AMD servers. The article reports the advantage falling to about 1.81× under a different memory allocation.
Why it matters
That distinction changes what the headline means. The comparison suggests Arm’s chip performance and power assumptions could let operators fit more systems, and therefore more workload capacity, into a fixed rack budget. It is not evidence that a pair of fully built racks were tested side by side.
Nor is “performance” one tidy number here. Results varied by workload: Arm led on some, while AMD led in MySQL and Milvus throughput. The tests also used different software environments, and Signal65 did not rerun the hardware independently in its own lab, according to the report.
Our read
This is a useful infrastructure comparison, but the rack figure deserves its footnotes in the same font as the headline. Arm’s approach asks a sensible question about useful work under limits on power and space; buyers still need comparable workloads, transparent assumptions and tests across complete systems before treating a modelled rack result as a purchasing answer.
What to watch
- Whether Arm, AMD or other chipmakers publish results using a shared agent-workload methodology.
- Whether later comparisons include measured performance from fully populated racks.
- How the results change with consistent software, memory allocations and system configurations.
Discussion spark: For AI infrastructure, should buyers put more weight on modelled performance per rack or on measured results from comparable systems?
Sources and evidence
- Arm says its AGI CPU delivers nearly 2X the performance per rack of AMD EPYC 9755 in new agentic AI CPU tests (19 October 2026, 11:00 UTC)
Independent WittyWires tracker for public updates about Arm. Not affiliated with or endorsed by Arm; this is not an official account.