Developer Tools

Arm AGI CPU

Arm's first-ever silicon product delivers 2x rack performance for agentic AI workloads, with Meta as lead partner.

Deep Dive

Arm has made a historic shift from IP licensor to silicon manufacturer with the announcement of the Arm AGI CPU. This new class of processor, built on the Arm Neoverse platform, is designed specifically to power the infrastructure for continuous, large-scale agentic AI systems where software agents coordinate tasks in real-time. The CPU is architected for rack-scale efficiency, with a reference 1OU, 2-node blade packing 272 cores. These blades can populate a standard 36kW rack with 30 units for a total of 8,160 cores, and a liquid-cooled 200kW design with Supermicro can house 336 CPUs for over 45,000 cores.

Arm claims the AGI CPU delivers more than 2x the performance per rack compared to the latest x86 systems, achieved through the fundamental advantages of the Arm architecture. Key to this is the class-leading memory bandwidth of the Neoverse V3 cores, which maintain performance under sustained load where x86 cores degrade, and high single-threaded performance that ensures 'every Arm thread does more work.' The commercial momentum is led by Meta, which is co-developing the CPU to optimize its gigawatt-scale infrastructure for its family of apps and to work alongside its custom MTIA accelerators. Planned deployments focus on accelerator management, agentic orchestration, and the densification of services needed for AI task scale-out.

Key Points
  • Arm's first-ever silicon product marks a historic shift from IP licensing to manufacturing, extending the Neoverse platform.
  • Designed for agentic AI, it scales to over 45,000 cores per rack and delivers 2x the performance of latest x86 systems.
  • Meta is the lead partner and customer, co-developing the CPU to work with its custom MTIA accelerators in gigawatt-scale infrastructure.

Why It Matters

It provides a foundational, high-performance CPU architecture purpose-built for the massive scale and continuous operation required by the next wave of agentic AI applications.