Ampere Targets More Efficient AI Compute

Ampere Computing’s AmpereOne architecture featuring a custom Arm-compatible CPU is designed to emphasize efficiency and performance per watt for cloud datacenter environments. The microarchitecture is scalable, with the initial AmpereOne product packing up to 192 CPU cores operating at up to 3.7 GHZ, connected through a custom mesh-based interconnect with a distributed coherence engine. Each core is equipped with a dedicated 2 MB L2 cache, and the initial product also features 64 MB of system-level cache. It implements innovations to improve datacenter virtualization and reduce bottlenecks such as highly accurate branch prediction and techniques to minimize memory disambiguation to prevent stalls.