Thrilled to unveil the CXL Process Near Memory (CXL-PNM) Disaggregation System at OCP Global Summit 2025! The Samsung Memory Architecture Group (MAG) is proud to have jointly developed this groundbreaking system with Marvell Technology and Liqid. This CXL-PNM system is a paradigm shift for new memory architecture, especially for Agentic AI. System Highlights: - Each CXL-PNM card delivers an impressive 200GB/s memory bandwidth and 4TB capacity. - It is powered by 16 ARM Neoverse V2 CPUs and the latest vector/matrix engines. - By linking up to 10 PNM cards, achieved 2TB/s memory bandwidth and 40TB capacity. This architecture allows Agentic AI workloads to offload TB-scale vector databases and KV caches—too large for GPU HBM—and directly process memory-bound operations like KNN/ ANN/ Attention. This effectively bypasses the bottleneck limitations of PCIe interconnects bandwidth. Our team presented the architecture in depth at the OCP Global Summit breakout session: - The Heterogenenous Memory Solutions for Agentic AI (by Jinin So): Showcasing significant performance and capacity enhancements via MRDIMM and CXL Memory Expander solutions. Introducing the PNM Engine directly into memory to alleviate congestion and stress on backend network(NVLink/UALink/SUE/CXL/PCIe) interconnects. - Software Enablement for PNM (by Wonjae Lee): Our progress and direction in developing the essential software to support the new PNM hardware architecture. - Advanced Memory Orchestration (by Han Deok Lee Taehoon Lee): Unveiling the Samsung DCMFM (Data-Centric Memory Fabric Manager) Software for efficient orchestration and management of these advanced memory resources. I will continue to focus on developing heterogeneous memory system technologies, including PIM/PNM, with the OCP Server/CMS/DCC FTI work streams to improve performance/$, and power efficiency for AI workloads. A huge thank you to our industry and academic partners for their invaluable support in developing this system: - Marvell Technology: Khurram Malik - Liqid: Sumit Puri - Academia: Professor Jung Ho Ahn(Seoul National University), Professor Nam Sung Kim (University of Illinois Urbana-Champaign), Professor Jungwook Choi(Hanyang University), Professor Ada Gavrilovska(Georgia Institute of Technology), - Samsung Electronics MSL : VP Sung-Wook Ryu, Mayank Saxena, Ramdas Kachare, Minyong Yoon We are committed to the continuous evolution of this system into 2026 with: - Explosive bandwidth and capacity improvements through LPDDR-based PIM integration. - SW Ecosystem expansion with Triton, OpenXLA, and oneAPI support. - Support for GPU backend networks (NVLink, UALink, SUE). We look forward to collaborating with you on the OCP platform to shape the future of AI infrastructure. Please reach out if you're interested in joining this exciting journey! #Samsung #PNM #CXL #AgenticAI #OCP2025 #DataCenter #AIInfrastructure #Marvell #Liqid #OpenComputeProject
-
-
-
-
-
+3