评估基于 CXL 的可组合内存解决方案：面向数据中心工作负载的总体拥有成本框架.pdf

《评估基于 CXL 的可组合内存解决方案：面向数据中心工作负载的总体拥有成本框架.pdf》由会员分享，可在线阅读，更多相关《评估基于 CXL 的可组合内存解决方案：面向数据中心工作负载的总体拥有成本框架.pdf（14页珍藏版）》请在三个皮匠报告上搜索。

1、Navneet Rao(Altera)Nilesh Shah(ZeroPoint)Contributors:Seema Mehta(Ampere),Mohamed El-Batal(Seagate),Khurram Malik(Marvell),Steve Scargall(Memverge),JM Hands(FarmGPU),Grant Mackey(JRL),Angelos Arelakis(ZeroPoint),Yiwei Yang(zett.ai),Vikrant Soman(Uber)Evaluating CXL-Based Composable Memory Solutions:

2、A TCO Framework for Datacenter Workloads2General Compute CMS HW solutionsCMS HW solutionsPooled memoryNative expansion(CXL PCIe devices)Data Streaming Kafka,Pinot,Athena,FlinkData Analytics SPARK,PrestoOnline StorageRedis,Cassandra,MySQLWeb AppsMicroservicesBaremetal hostsLinux,WindowsContainers,VMs

3、K8s,KVM,VmwareFabric managers Cluster/Resource managers Generative AI&Traditional MLCMS HW solutionsCMS HW solutionsShared memory(pooled and shared)Native expansion(HBM?)GenAI Training HuggingFace,LLMsGenAI ServingHuggingFace,RAG,LLMs,fine tuning Traditional ML inferenceXGBoost,DLRMTraditional ML tr

4、aining XGBoost,DLRMBaremetal hostsLinux,WindowsFabric managers GenAI orchestration K8s,RayBenchmarkingFBGEMM.MLPerf,vectorDB,GenAI benchmarks?Traditional ML orchestration K8s,Ray,SPARKNative expansion(CXL devices)Deployment frameworks(Tensorflow,Pytorch,Huggingface Xmers)NewOCP CMS Workload subgroup

5、:2025 charterDemonstrate price/performance benefits for datacenter workloads using CMS BenchmarkingSPECCPU,memtier,cassandra-stress etc32026 focus areasTCO model for General compute to demonstrate value proposition of CMS Benchmark development,execution and submission to OCP CMS repo Lift-&-shift do

6、ne to CMSRepo(from Memverge github)Develop Gen AI/Traditional ML charterWorkload characterization,feasibility study,prototyping,benchmarking etcGeneral compute+CMS pooled solution-demonstrate improved memory utilization/efficiency by reducing stranded memory+CMS native expansion Continue workload as