Arm detailed an upcoming AGI server CPU featuring two 70-core N3P chiplets during the Hot Chips 2026 conference [1].
The announcement signals Arm's intent to capture a larger share of the server market by providing the hardware infrastructure necessary for artificial general intelligence. By showcasing these designs, the company aims to attract server-market customers and demonstrate its near-term CPU roadmap [1], [2].
The processor architecture utilizes a chiplet-based design. The CPU consists of two 70-core N3P chiplets, allowing for a total core count of up to 136 cores [1]. This modular approach allows for greater flexibility in manufacturing and scaling.
To handle the massive data throughput required for AGI workloads, Arm integrated a UCIe fabric link with a bandwidth of 2 TB/s [1]. This high-speed interconnect is critical for maintaining communication efficiency between the chiplets.
Memory management is handled by a 12-channel memory controller [1]. This configuration is designed to reduce bottlenecks when the CPU accesses large datasets, a common requirement for large-scale AI models.
Despite the technical specifications provided, Arm did not disclose specific performance figures for the CPU [1]. The company focused instead on the architectural capabilities and the physical specifications of the N3P process.
“The CPU consists of two 70-core N3P chiplets, allowing for a total core count of up to 136 cores”
The move toward chiplet-based architectures with high-bandwidth interconnects like UCIe indicates that Arm is prioritizing scalability and data movement over single-core clock speeds. By targeting AGI workloads, Arm is positioning itself to compete directly with established server chip giants, focusing on the specific memory and throughput demands of next-generation AI training and inference.
