Arm Unveils AGI CPU

Artificial intelligence chips do not just train models; they also run the steady back-office work that keeps answers flowing. Arm said on March 24 it is now selling its own data-center processor, the Arm AGI CPU, to handle that layer of inference and agent software. (arm.com) That is a break from Arm’s business for most of its 35-year history. The company has usually licensed chip designs and instruction sets to others, but it said the Arm AGI CPU is its first Arm-designed production silicon product. (arm.com) The chip is aimed at data centers running “agentic” systems, meaning software that can plan, call tools, move data, and manage many parallel tasks after a model generates tokens. Arm said those workloads are pushing demand for more central processing unit capacity per gigawatt of data-center power. (arm.com; arm.com) Arm’s pitch is density: more useful compute in the same rack, power, and cooling budget. In its launch materials, the company said the processor delivers more than 2x performance per rack versus comparable x86 systems and can scale to 8,160 cores in a standard 36-kilowatt air-cooled rack. (arm.com; arm.com) The basic idea is that graphics processors handle much of model training and the first burst of inference, while central processors coordinate memory, networking, storage, and follow-on tasks. HPCwire reported that Arm is targeting the “decode” stage of inference, where systems may run code, query databases, and manage feedback loops around model output. (hpcwire.com) Arm’s product brief says the top version has up to 136 Neoverse V3 cores, up to 3.2 gigahertz all-core speed, boost up to 3.7 gigahertz, 96 lanes of Peripheral Component Interconnect Express Gen 6, and support for Compute Express Link 3.0. Arm also lists up to 6 gigabytes per second of memory bandwidth per core and a 300-watt thermal design power. (arm.com) Meta was the lead development partner, according to Arm, and OpenAI is among the companies publicly endorsing the chip. Arm’s ecosystem page also lists Cerebras, Cloudflare, Lenovo, Positron, and server makers including ASRock Rack as launch partners. (arm.com; arm.com) Arm is also shipping a server blueprint alongside the processor. Its developer blog describes a 1U dual-node reference server that puts two compute nodes in one chassis at about 1,100 watts nominal power, a design meant to raise node density per rack unit. (developer.arm.com) The move places Arm closer to customers that already build their own Arm-based chips, including large cloud companies. Arm said it will keep offering intellectual property and Compute Subsystems alongside its new silicon, framing the chip as another option rather than a replacement for its licensing model. (arm.com) What happens next is less about a single chip launch than adoption in live fleets. Arm has said early systems are already in the market, and its first test will be whether companies buying inference capacity want Arm not only as the architecture underneath their servers, but as the chip supplier too. (arm.com; trendforce.com)