Compute Express Link

Overview

Compute Express Link (CXL) is an open interconnect standard built on top of PCIe that enables memory expansion, memory pooling, and near-data computation for AI and high-performance computing workloads. CXL addresses the capacity wall that emerges when large AI models exceed the memory addressable by direct CPU/GPU-attached DRAM: by attaching fabric-connected memory modules via CXL, servers can expand usable memory without hardware replacement.

Three capabilities define CXL's value for AI infrastructure:

1. Memory expansion — CXL memory modules (CXL Type 3) connect as cache-coherent memory, allowing LLMs to run on existing server hardware by expanding available capacity.
2. Memory pooling — Starting with CXL 3.0, real-time memory sharing across multiple servers enables dynamic allocation, improving scalability and reducing overall system build cost.
3. Near-data compute — CXL Memory with Compute (CMM-DC) embeds accelerator logic in the CXL controller, performing computation close to the data and minimizing data movement across the PCIe bus.

AI workloads' relative insensitivity to memory latency — compared with transactional workloads — makes long-latency fabric-attached CXL memory a viable complement to high-bandwidth near-compute solutions like HBM.