Zhi Tong Cai Jing·2025-10-29 02:29Core Insights - The report from GF Securities highlights the synergy between MRDIMM and CXL in enhancing AI server memory supply and elastic expansion, addressing the challenges of high cost and limited capacity of HBM, diverse memory demands, and CPU memory expansion bottlenecks [1] Group 1: MRDIMM and CXL Benefits - MRDIMM and CXL create a "near-end high bandwidth + far-end large capacity" layered collaboration to increase AI server memory supply and elastic expansion at a lower TCO [1] - MRDIMM provides deterministic gains in KVCache scenarios, with higher concurrency, longer context, and lower end-to-end latency, significantly optimizing CPU-GPU memory orchestration [2] - MRDIMM Gen2 supports speeds up to 12800MT/s, offering a 2.3 times bandwidth increase over DDR5 RDIMM under AI loads, which reduces KVCache read/write latency and supports high-throughput inference [2] Group 2: CXL Advantages - CXL 3.1 significantly enhances performance for KVCache, particularly for high concurrency and ultra-long context loads [3] - CXL enables memory pooling and expansion, allowing KVCache to be elastically offloaded from expensive GPU memory to CXL devices, expanding effective capacity to TB levels without increasing GPU costs [3] - The decoupled KVCache architecture allows for a 30% increase in batch size and a 87% reduction in GPU demand, with a 7.5 times increase in GPU utilization during the prefill phase [3]