Loading...
Loading...
Arm is shifting from its traditional IP-licensing model toward selling its own datacenter silicon, unveiling a 136-core “AGI CPU” aimed at powering agentic AI workloads that coordinate GPUs and accelerators rather than running models directly. Built on TSMC’s 3nm process and targeting production availability in H2 2026, the dual-die Neoverse V3 design emphasizes deterministic single-thread performance and efficiency, while offering modern I/O like PCIe 6.0 and CXL 3.0. Meta is positioned as an early customer, with OpenAI and others sampling, signaling rising CPU demand in AI infrastructure—and potential tension with Arm’s existing chipmaking partners.
Arm unveiled the ARM AGI CPU on March 24, 2026 — its first production silicon aimed at large-scale AI infrastructure, delivering up to 136 Neoverse V3 cores, Armv9.2 with bfloat16/INT8 AI instructions, 3.7GHz boost, 96 PCIe Gen6 lanes with CXL 3.0, and a 3nm process. The dual-chiplet design supports up to 6 TB DDR5-8800 across 12 channels and up to 420W TDP. Arm offers three SKUs: a 136-core flagship (SP113012), a 128-core TCO-optimized variant (SP113012S), and a 64-core memory-bandwidth-optimized model (SP113012A). Reference servers include a 10U, 2-node air-cooled blade packing 272 cores per blade and a Supermicro liquid-cooled 200kW chassis that can scale to over 45,000 cores. This targets hyperscalers and AI datacenters seeking dense, Arm-native compute for agentic workloads.
Arm unveiled the ARM AGI CPU at its March 24, 2026 Arm Everywhere keynote — its first production silicon aimed at large-scale AI infrastructure. The chip packs up to 136 Neoverse V3 cores, Armv9.2 with bfloat16/INT8 AI instructions, up to 3.7GHz boost, 96 lanes of PCIe Gen6 with CXL 3.0, a 3nm process, up to 420W TDP, and support for up to 6TB DDR5-8800 via a dual-chiplet design. Three SKUs target different trade-offs: a 136-core flagship (SP113012), a 128-core TCO-optimized part (SP113012S), and a 64-core bandwidth-optimized SKU (SP113012A). Reference servers scale to 8,160 cores in air-cooled racks or over 45,000 cores in a Supermicro liquid-cooled design, signaling Arm’s push into dense, agentic AI datacenter deployments.
Arm announced it will produce its own semiconductor — the Arm AGI CPU — shifting from its decades-old licensing model to selling in-house chips for AI-optimized servers. Fabricated by TSMC on a 3nm process, Arm bills the AGI CPU as an energy-efficient, agentic workload CPU with superior performance-per-watt versus x86 rivals; it aims for full production availability in H2 2026. Early customers and sample recipients include Meta, OpenAI, SAP, Cerebras, Cloudflare, SK Telecom and Rebellions. Arm showcased industry endorsements from Nvidia, Amazon and Google executives (no purchase commitments), underscoring broad interest but also raising potential partner tension as Arm moves into the maker role once held by Intel, AMD and others.
Arm announced its first in-house datacenter processor, the 136-core “AGI CPU,” a dual-die Arm Neoverse V3 design aimed at agentic AI workloads rather than running models directly. Fabricated on TSMC 3 nm, the 300 W chip clocks to 3.7 GHz, offers 2 MB L2 per core plus 128 MB system cache, 12-channel DDR5 (up to 8800 MT/s), 96 lanes PCIe 6.0 and CXL 3.0. Arm intentionally omitted legacy features and SMT to favor deterministic, single-thread performance for agents that orchestrate GPUs/AI accelerators. Meta is a launch customer and Arm validated high-density OCP rack designs (air- and liquid-cooled) to scale deployments. The move signals Arm shifting from IP licensor to full-platform silicon provider to capture rising CPU demand from agent-driven AI infrastructure.