Arm Claims 2x Rack-Level Performance Over AMD With Its First-Ever In-House CPU

Quick Facts

• Arm's AGI CPU delivered 1.99x the modeled agentic work per rack of AMD's EPYC 9755 across seven workloads in a Signal65 benchmark study released at OCP Global Summit 2026.
• The chip marks the first time in Arm's 35-year history that the company has shipped its own production silicon, moving beyond its traditional IP licensing model.
• The rack-level advantage is partly structural: a 24 kW rack fits 21 to 23 AGI CPU servers versus 13 dual-socket AMD systems, yielding up to 5,888 Arm cores against 3,328 for AMD.

Arm unveiled a benchmark study at OCP Global Summit 2026 in San Jose crediting its new AGI CPU with nearly twice the modeled agentic work per rack of AMD's EPYC 9755, a claim that carries weight for data center operators but comes with notable caveats.

The study, authored by Russ Fellows, VP of Labs at Signal65, evaluated seven workloads across tooling, agentic workflows, and reasoning tasks. Average gains were 1.7x for tooling workloads like MySQL and NGINX, 2.0x for agentic orchestration including Terminal-Bench 2 and SWE-bench replay, and 2.2x for reasoning and planning tasks like Milvus vector search and vLLM CPU overhead.

A Historic Business Model Shift

Arm announced the AGI CPU on March 24, 2026, extending its Neoverse platform into production silicon for the first time. The company has spent more than three decades licensing chip architecture to partners including Apple, Qualcomm, and Amazon. Shipping its own processor puts Arm in direct competition with some of those same customers.

CEO Rene Haas framed the move in terms of the shifting workload mix. "AI has fundamentally redefined how computing is built and deployed," he said. "Agentic computing is accelerating that change."

What the Chip Actually Is

Arm detailed the AGI CPU's architecture at Hot Chips 2026. The chip uses a dual-chiplet design built on TSMC's N3P process, with configurations of 64, 128, or 136 Neoverse V3 cores and a 300W TDP. Each chiplet holds 70 cores, four of which are redundant for yield, connected by a 2 TB/s UCIe link.

Memory bandwidth reaches 844.8 GB/s via DDR5-8800, with sub-100ns DRAM latency. The chip supports 96 PCIe 6.0 lanes and CXL 3.0. At 300W for 136 cores, Arm claims 2.2 watts per core, compared with 3.9 watts per core for Intel's 128-core Xeon 6 at 500W.

How the Rack Math Works

The benchmark's headline number is partly an engineering story about power density. An Arm AGI CPU server draws 950 to 1,050 watts. A dual-socket AMD EPYC 9755 system draws 1,700 watts. Inside a 24 kW rack, that difference allows 21 to 23 Arm servers versus 13 AMD servers.

That configuration delivers between 5,712 and 5,888 physical cores for Arm, against 3,328 for AMD, before any per-core performance difference is applied. Arm and Signal65 introduced an "agentic work per rack" metric specifically to capture this dynamic, positioning it against simpler agents-per-rack marketing claims from competitors.

Sandbox creation latency tests, using the Sandbench tool against AMD's EPYC 9965, showed Arm completing cold starts 45% faster, warm starts 27% faster, and worst-case scenarios under fourfold oversubscription 41% faster. The report also estimated database and vector-search operations could require 30% to 35% fewer racks for equivalent processing loads.

What Is Still Missing

The study was conducted by Signal65 with Arm's cooperation and relies partly on power-based modeling, with testing on final production hardware still outstanding. Signal65 noted it plans follow-on testing across a broader workload range, including its own PINNACLE CPU-centric benchmarks.

Arm has not published conventional results such as SPEC CPU2017, SPECrate, or direct socket-to-socket comparisons against current AMD EPYC or Intel Xeon processors under standard server workloads. The primary performance claim remains a rack-level estimate, not a socket-level measurement.

Arm developed the benchmarking framework with support from Fujitsu, Meta, and OpenAI. The company is targeting $2 billion in chip demand by 2027 to 2028 and $15 billion in chip sales by 2031, according to public disclosures. Whether the AGI CPU's rack efficiency translates into data center contract wins against entrenched x86 suppliers depends on benchmarks that have not yet been published.