136-Core Processor for Data Center AI Arm Presents First Self-Produced AI Processor

Related Vendors

FAULHABER Drive Systems

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

With the AGI CPU, Arm is moving away from the pure IP licensing model and selling its own chips for data centers. The 136-core processor is aimed at AI infrastructures for agent-based systems.

With the AGI CPU, Arm has introduced its own processor for data centers for the first time, thereby initiating a strategic change of course. Until now, the company has mainly been active as a provider of chip designs and architecture IP, which partners such as Qualcomm, Nvidia and other manufacturers license.

The announced processor is designed for use in AI infrastructures, in particular for so-called agentic systems. This refers to applications that perform tasks largely independently, write code or coordinate processes and require more general computing power in addition to accelerators.

According to Arm, the CPU primarily performs orchestration tasks in such environments. It should control data movements, connect accelerators and cover the general compute requirements, while the actual training or inference continues to run mainly on GPUs or specialized AI components.

With this step, Arm is adding a third pillar to its existing business model. In addition to traditional IP licensing and compute subsystem offerings, the company now also sells ready-made silicon that customers can integrate directly into their infrastructure.

From Pure IP Provider to Chip Supplier

Technically, the AGI CPU is based on up to 136 Neoverse V3 cores. The processor is manufactured at TSMC using the 3-nanometer process, consists of two dies and is designed for a power consumption of 300 watts. The clock frequency is up to 3.7 GHz, the base clock is 3.2 GHz.

Arm relies on twelve DDR5 channels with data rates of up to 8800 MT/s for the memory connection. In addition, there are 96 PCIe 6.0 lanes and support for CXL 3.0. It is striking that Arm has dispensed with additional accelerator functions in the chip and instead focuses on a CPU design tailored to the target workloads.

Arm is also aiming for high packing density when it comes to system configurations. These include air-cooled 36 kW racks with 30 blades and a total of 8,160 cores as well as liquid-cooled 200 kW configurations with more than 45,000 cores. The company is thus positioning itself for large-scale AI and data center environments.

The lead partner of the project is the Facebook group Meta, which plans to use the processor on a larger scale this year, according to Arm. According to the company, both companies worked together on the design. Meta plans to use it in combination with its own accelerators.

Aiming for High Growth Targets

According to Arm, early customers and partners also include OpenAI, Cloudflare, SAP, SK Telecom, Cerebras, F5, Lenovo, Quanta and other companies. Arm sees the area of application not only in agent-based AI workloads, but also as a control and head node for accelerators as well as in classic infrastructure tasks such as network or storage workloads.

For Arm, entering the market with its own silicon is also financially ambitious. CEO Rene Haas put the medium-term sales potential of the new processor at around 15 billion US dollars per year within around five years. At the same time, the traditional IP business should also continue to grow.

Arm is thus attempting to benefit from the expansion of the AI infrastructure on several levels. The company not only wants to earn money from the licensing of future designs, but also wants to participate directly in the sale of processors and complete system platforms.

6 Gigawatts of GPU Capacity

Meta and AMD Agree on Delivery of AI Chips Worth 60 Billion US Dollars

Whether this step will create long-term tensions with existing partners remains to be seen. With the AGI CPU, Arm is now entering markets in which the company's customers and licensees are already active with their own server processors.

In the short term, the AGI CPU should primarily be seen as a signal: Arm no longer wants to position itself in the data center merely as a technology provider, but as an independent provider of products for AI infrastructures. For the chip industry, this is a relevant change in strategy with a potential impact on the distribution of roles, partnerships and competition in the data center market.(sg)