UALink

Overview

UALink is the only open standard specifically designed and purpose-built for AI GPU scale-up networking, developed by the UALink Consortium with Intel and Astera Labs as board-level promoter members. It delivers 200 Gbps per direction in a x4 configuration (800 Gbps full-duplex), with very low latency and memory-semantic load/store/atomic operations, scaling to 1,024 GPUs in a single scale-up domain.

Intel positions UALink as more efficient than proprietary scale-up interconnects because it carries no legacy overhead from prior use cases—its architecture was designed specifically for AI workload communication patterns. The consortium's active roadmap as of October 2025 includes in-network collectives, a UCIe-based IO chiplet integration, confidential compute support, and software ecosystem manageability.