Arm Holdings plc today debuted a new central processing unit, the AGI CPU, that’s optimized to power artificial intelligence clusters.

The company says the chip provides more than twice as much performance per server rack as Intel Corp. silicon. Additionally, Arm claims, the AGI CPU can help data center operators reduce hardware costs. The company expects the chip to provide savings of up to $10 billion per gigawatt of data center capacity.

Arm is best known for designing CPU cores, the parts of a processor that carry out calculations. It also makes many of the other components necessary to build a chip. The company sells a mesh interconnect, a kind of miniature network that can link together multiple cores into a functioning CPU. The module is available alongside subsystems that perform supporting tasks such as regulating voltage levels.

AMD licenses its CPU component designs to chipmakers that assemble them into processors. Until now, however, the company didn’t sell a complete processor of its own. The launch of the AGI CPU marks Arm’s entry into that market.

“With the expansion into delivering production silicon with our Arm AGI CPU, we are giving partners more choices all built on Arm’s foundation of high-performance, power-efficient computing, to support agentic AI infrastructure at global scale,” said Arm Chief Executive Officer Rene Haas.

The AGI CPU comprises two dies made using a three-nanometer manufacturing process. The dies host up to 136 of Arm’s latest Neoverse V3 server CPU core design. The core’s instruction set, the language in which it expresses computations, includes extensions specifically optimized for AI workloads.

Neoverse V3 also ships with multiple data protection features. A technology called RME enables it to run workloads in isolated, encryption sections of a server’s memory to block hacking attempts. The core can also automatically fix data errors that emerge in its cache. Each of the AGI CPU’s cores includes 2 megabytes of L2 cache.

Arm optimized the chip for sustained performance, which means that it can maintain a consistent speed for extended time periods. Each of the AGI CPU’s cores runs a single thread at a frequency of up to 3.7 gigahertz.

The company has released a reference server design for data center suppliers that plan to integrate the processor into their products. It’s an air-cooled blade, or compact server, with two AGI CPUs. Data center operators can install up to 30 such blades in a standard air-cooled 36kW server rack for a total of 8,160 cores.

Several hardware makers including Lenovo Group Ltd. are already shipping servers based on the AGI CPU. Arm says its initial customer roster includes OpenAI Group PBC, Cloudflare Inc., Cerebras Systems Inc., SAP SE, Meta Platforms Inc. and several other companies. Facebook parent Meta is the AGI GPU’s lead customer.

Arm’s partnership with Meta also extends to other areas. The companies optimized the AGI GPU to work with the social network’s custom MTIA family of AI chips. Furthermore, Meta and Arm plan to co-develop several future generations of AI-optimized CPUs.

Image: Arm