当前位置:首页 > 报告详情

利用智能体人工智能和以内存为中心的计算实现异构内存机遇.pdf

上传人: 明**** 编号:1011773 2025-12-21 25页 1.99MB

1、Jinin So,Senior Director&System ArchitectMemory Division,Samsung Electronics Khurram Malik,Senior DirectorMarvell TechnologyHeterogeneous Memory Opportunity with Agentic AI and Memory Centric ComputingHeterogeneous Memory Opportunity w/Agentic AI and Memory Centric ComputingJinin So,Senior Direct&Sy

2、stem Architect,Samsung ElectronicsKhurram Malik,Senior Director,Marvell TechnologySERVER:COMPOSABLE MEMORY SYSTEMS(CMS)Agentic AI is an intelligent system which operates autonomously,making adaptive decisions and taking actionsto achieve specific goals with minimal human interventionLLMs play a cruc

3、ial role with natural language understanding/planning/reasoning capability Some popular examples:Agentic RAG(ChatGPT,Perplexity),Coding Agent(Claude Code,Gemini CLI)Agentic AIFinal answerAnswerAnswer to UserAnswerAnswer to UserReflectionIs the responseRelevant?(multiple)ToolsUser QueryReflectionIs t

4、he responseRelevant?ReflectionIs the responseRelevant?ToolsToolsAnswerPlanningMake plan&Tool callPlanningMake plan&Tool callTool Call Agentic AI based on LLM requires significant memory capacity and bandwidth relying on external databases(Working)A space where agent temporarily stores and processes

5、information current tasks or reasoning(Procedural)Implicit knowledge stored in LLM weights,(Semantic)Agents knowledge about the world and itself(Episodic)Experiences from earlier decisions(e.g.training input-output pairs,history event flows,and game trajectories)Agentic AI:Memory ViewParseParseRetri

6、evalRetrievalProcedural MemoryPromptPromptLearningLearningRetrievalRetrievalLearningLearningSemantic MemoryRetrievalRetrievalLearningLearningEpisodicMemoryDecision ProcedureLLM with Reasoning(Working Memory)Step1Step2Step3OutputActionsActionsObservationObservationParseRetrievalProcedural MemoryProce

word格式文档无特别注明外均可编辑修改,预览文件经过压缩,下载原文更清晰!
三个皮匠报告文库所有资源均是客户上传分享,仅供网友学习交流,未经上传用户书面授权,请勿作商用。
根据报告的内容,全文主要探讨了Agentic AI和Memory Centric Computing在异构内存系统中的应用,以及相关技术挑战和解决方案。以下是关键点: 1. Agentic AI需要大量内存和带宽,依赖于外部数据库。 2. LLMs在自然语言理解、规划、推理中起关键作用。 3. Procedural Memory由LLM权重构成,存储执行任务的方法。 4. Semantic Memory存储关于世界的通用知识,通常以向量数据库形式实现。 5. Episodic Memory存储先前决策的经验。 6. LLMs需要更多内存带宽/容量以满足服务延迟协议(SLA)。 7. HBM提供高带宽,但容量有限且成本高。 8. CMM-D和MRDIMM提供高容量和带宽,但成本高且带宽有限。 9. PNM/PIM技术可支持大规模向量数据库和高效KV-Cache管理。 10. CXL-PNM和CXL-PIM技术可提高语义记忆和工作记忆的性能。
揭秘HBM与CMM-D" "如何突破LLM带宽瓶颈?" 语义记忆搜索加速新篇章"
客服
商务合作
小程序
服务号
折叠