Samsung launches CXL memory modules to boost AI inference performance

Samsung Electronics unveiled a CXL (Compute Express Link) based DRAM memory module, the CMM‑D, aimed at relieving AI inference bottlenecks in large‑language‑model workloads. In tests using Nvidia RTX 6000 GPU servers, a 1 TB CXL memory pool delivered performance comparable to DRAM in a single‑GPU setup and retained about 92 % of DRAM performance when eight GPUs were used together. The larger pool also prevented the performance drop seen with KV‑cache‑heavy 512 GB DRAM configurations, keeping token‑per‑second rates stable.

The company positions CXL pooling as a “next‑generation HBM” solution that can expand server memory capacity while keeping latency low, and it plans to advance CXL 3.0‑compatible production. Market analysts project the global CXL market to grow from roughly $2.1 billion in 2023 to about $16 billion by 2028, driven by data‑center demand for scalable AI hardware. Samsung notes that broader ecosystem adoption—including CPU support for CXL—will be critical for wider deployment.