Arm AGI: A New Processor for AI Servers
Arm has introduced the AGI server processor featuring up to 136 Neoverse V3 cores and support for up to 6 TB of memory, designed specifically for artificial intelligence workloads. Commercial shipments are expected to begin in the coming months.
Crius
New Arm AGI Server Processor: Key Features
Arm has introduced the AGI server processor, designed specifically for artificial intelligence workloads. The model is equipped with up to 136 Neoverse V3 cores, supports up to 6 TB of RAM, and delivers memory bandwidth of up to 844.8 GB/s. Commercial shipments of the processor are expected to begin in the coming months.
Architecture and Performance
The Arm AGI processor is manufactured using TSMC’s N3P process and utilizes two nearly autonomous chiplets. Configurations are available with 64, 128, or 136 Neoverse V3 cores, operating at frequencies from 2.8 to 3.7 GHz. Each chiplet contains 70 physical cores, with four cores that can be disabled to improve yield in the 136-core version. Every core features two 128-bit vector engines and 2 MB of L2 cache, while the total system-level cache reaches 272 MB.
Memory and Bandwidth
Each chiplet integrates compute, memory, and I/O subsystems. The chiplets are connected via a UCIe interface at 32 GT/s, providing a combined bandwidth of up to 2 TB/s. Unlike some other solutions, each chiplet includes its own CPU cores, memory controllers, and I/O components, allowing local memory access without crossing chiplet boundaries and reducing latency. Access time to local DRAM can be less than 100 ns.
Memory is one of the processor’s key advantages. Each chiplet is equipped with a six-channel DDR5 controller, delivering up to 844.8 GB/s of total memory bandwidth. DDR5-8800 is supported, with up to 6 TB of memory per socket. Mechanisms for memory traffic prioritization and management of contention between CPU cores and I/O devices are implemented.
Scalability and Power Consumption
To expand AGI’s capabilities, the processor offers 96 PCIe 6.0 lanes with CXL 3.0 support, enabling memory expansion, as well as four PCIe 4.0 lanes and interfaces for I3C, I2C, and SPI. The processor’s TDP is 300 W.
Applications and Market Position
Arm positions AGI for use in AI servers and agent-based AI systems where high memory bandwidth and minimal latency are critical. According to the company’s internal estimates, servers based on the new processor can deliver up to twice the rack-level performance compared to systems using modern x86 processors. However, these figures are based on Arm’s own assessments.
Outlook
With the launch of AGI, Arm is strengthening its presence in the server market, which has traditionally been dominated by x86 processors. Commercial availability of Arm AGI is expected to begin in the coming months.
