下一代多头可组合存储器设备 (MHCMA).pdf

1、Mohamad El-Batal Seagate Office of the CTOHongjian Fan Seagate Research GroupMulti-Headed Composable Memory Appliance(MHCMA)Next-GenMulti-Headed Composable Memory Appliance(MHCMA)Next-GenMohamad El-Batal Seagate Office of the CTOHongjian Fan Seagate Research GroupSERVER:COMPOSABLE MEMORY SYSTEMS(CMS

2、)Understanding the current CXL ecosystem challengesDiscuss the current CMA/CFM architecture limitationsExplore enhancements to make a CXL Multi-Headed CMA viableFuture MHCMA TCO analysis and opportunitiesMHCMA in Memory Fabric OrchestrationCall to ActionOutlineCXL Ecosystem 2024/2025 ChallengesCXL p

3、rotocol was created to solve two datacenter problems:Memory Wall Memory scaling is 1/10thCPU Core scalingStranded Memory Oversized&Underutilized Server Locked-in Memory resourcesInitial 2024/2025 CXL deployment with memory-expansion&pooling was a disappointmentMemory-Expansion alone Attempted to sol

4、ve the Memory-Wall problem,but did nothing for the Stranded-Memory issue Existing 16GB&32GB DDR4 DIMMs inventory redeployment couldnt TCO justify the higher-cost of expansion/pooling hardware12x DIMMs per card plus 2:1 HW Memory-Compression are a minimum to somewhat satisfy the CXL expansion TCO mod

5、elThe Contributed CMA&CFM Solution was only a BeginningWhat Should be the Next MHCMA&CFM ArchtiectureIndustry looking for a TCO optimized MHCMA&CFM Memory Pooling/Sharing Composable CXL Memory-Pooling/Sharing with 8-Port MHCMA/CFM to optimize Node Memory costCost optimized 8-Blade appliance with12 t

6、o 16 DDR4/DDR5 DIMMs per bladeLess than one NUMA Hop Average Typical CXL cache-line Latency with Minimal tail latencyHigher-BW Memory Pooling CXL3.x Gen6(Cloud)and Sharing CXL4.x Gen7 for GPU-nodesIntroduce Optical Linear-drive QSFP-DD/OSFP-XD cabling to reduce copper cabling complexity/clutterLever