At the OCP Global Summit in San Jose, Arm promoted an independent analysis of its AGI CPU. Signal65 wrote the analysis, and Arm commissioned it. It pits Arm's first production server silicon against AMD's fifth-generation EPYC 9755 (Zen 5, "Turin"). Server CPUs are not Windows desktop silicon. Even so, they decide what Windows Server, Azure-adjacent and Linux estates run on. Capacity planners and infrastructure admins will care about the numbers, as long as they read them carefully.
The headline number
Signal65 reports that across seven workloads the AGI CPU delivers an average of 1.99x the modeled performance per rack of the EPYC 9755, with individual workloads ranging from 1.43x to 3.19x. The seven CPU workloads span agent execution, tooling, retrieval and model-serving support. The study mixes common infrastructure components with newer agentic benchmarks, including Terminal-Bench 2 and SWE-bench. Coverage of the report names the full set as MySQL, NGINX, Terminal-Bench 2, SWE-bench replay, Milvus vector search, vLLM CPU overhead and fastText.
By category, the AGI CPU averages 1.7x for tooling, 2.0x for end-to-end agentic workflows and 2.2x for reasoning and planning in modeled rack-level performance.
The key word is "modeled". The model places 21 to 23 AGI CPU servers in a 24 kW rack, against 13 EPYC 9755 servers. That works out to 5,712 to 5,888 physical Arm cores against 3,328 AMD cores. Nobody wired up a full rack and ran the benchmarks on it.
Where the 2x really comes from
The power assumptions carry much of the result. The assumed power per two-socket server is 950–1,050 watts for Arm and 1,700 watts for AMD. Those are rated figures, not measured draw under load. More Arm servers fit under the same power cap, and that density is what lifts the rack score. Signal65 itself calls rack-level density the key driver.
Futurum's analysis of the report fills in the per-chip picture. Futurum says AGI CPU averages 1.20x the EPYC 9755 per chip across the seven workloads. It also says AMD wins two outright at socket level: MySQL at 0.88x and Milvus at 0.95x. Arm's own power-based density factor, which Futurum puts at 1.62x to 1.77x, accounts for the rest of the jump to roughly 2x.
So the claim that the AGI CPU does twice the work per CPU is not what the report shows. The defensible reading is about twice the modeled work per rack under a fixed power budget.
The memory assumption
The report's sensitivity analysis sizes memory at 4 GB per hardware thread. That gives the SMT-enabled EPYC about 2 TB per server against 1 TB for the AGI CPU. Sizing per physical core instead would cut the seven-workload average from 1.99x to about 1.81x. That is still a large lead, but a visibly smaller one. Memory capacity, cost and power all shift the real-world math.
Early hardware and unattributed gains
Signal65 notes that the rack results are modeled rather than measured. It also says the per-chip results come from early AGI CPU systems, and that the tests make no attribution to SMT, cache or instruction set. In other words, the study doesn't say why the Arm chip does better in each case. Signal65 also says it plans follow-on testing on final production hardware, with more workloads and more competitive comparisons.
Futurum offers a fair verdict: the transparency is the strongest part of the study, and buyers should check whether Arm's rated power figures and the 4 GB-per-thread memory assumption match their own deployment math.
A separate test: sandbox startup
Don't fold this into the rack figure. Signal65 compared the AGI CPU with a two-socket AMD EPYC 9965 on Sandbench, across the runc, Firecracker and gVisor sandbox runtimes. The result was 45% less time for cold starts, 27% less for warm starts and 41% less in the worst case under 4x oversubscription. That measures how quickly isolated environments start up for agents. It does not mean every agent task runs 41–45% faster. Note too that this test used a different AMD part, the EPYC 9965, from the rack comparison.
What's missing: Venice
Wccftech points out that the comparison leaves out AMD's Zen 6-based EPYC Venice, which the outlet says brings large single-core gains. None of the sources I reviewed contain an AGI CPU versus Venice result. A fair reading is that the question is open, not that either chip would win. The tested parts are Turin, the 9755 for the rack comparison and the 9965 for sandboxes.
Why admins should care
- Power is the real constraint. Signal65 frames the rack as the critical unit of performance, since power and space constrain AI infrastructure growth. Any comparison that uses rated rather than measured power deserves a sanity check against your own facility.
- Vendor-funded research is still vendor-funded. Arm commissioned this work. The data may well be sound, and the workloads were chosen by people with an interest in the outcome.
- Software compatibility still decides. Arm servers already have footholds at the big clouds. Signal65 lists AWS Graviton, Google Axion, Microsoft Azure Cobalt and NVIDIA Vera among Arm-based infrastructure efforts. For on-premises shops running x86-only stacks, the migration cost does not appear in any rack-density chart.
- Wait for production silicon. Signal65 says it will retest on final hardware. Hold procurement decisions until those numbers exist.
Bottom line
Arm has produced a credible, openly caveated argument that density and power efficiency can give agentic workloads about twice the modeled work per rack. The same data shows AMD still ahead on MySQL and Milvus per socket, and a rated-power model that does the heavy lifting. The memory assumption trims the lead to roughly 1.81x. Zen 6 Venice remains untested. Treat the 2x as an opening bid, not a verdict.