以开放的速度扩展人工智能：从硅到系统.pdf

《以开放的速度扩展人工智能：从硅到系统.pdf》由会员分享，可在线阅读，更多相关《以开放的速度扩展人工智能：从硅到系统.pdf（22页珍藏版）》请在三个皮匠报告上搜索。

1、Sachin Katti Senior Vice President Chief Technology and AI OfficerIntelScaling AI at the Speed of Openness:From Silicon to SystemsCommitmentCommitmentto sustainable,flexible,and scalable tech infrastructure.Impact in collaboration with the Open Compute ProjectChoiceChoiceBreaking free from proprieta

2、ry constraints90 C90 ContributionsontributionsFrom OCP-NIC to DC-MHS to Open Systems for AIStrategic PriorityAI as aMassive transformation in AIRedefining every layer of the stack deployed from AI PC,to edge,to data centerCPU relevance in AI ecosystemOpen heterogeneous strategy to deliver systems,sy

3、stems,software&GPUs80%of AI compute cycles in inference by 2028*Physical AI Real-time inferencing for physical world interaction20152020202520302035Agentic AI Billions of daily inferences across thousands of modelsAgentic AI is the Means to Unlocking EnterpriseValuein AIDiverse workload requirements

4、Throughput,low latency,perf/$/W and system-level tuning required*Source:Gartner,published 08/2024140 xper monthCritical to optimize system-level tokens per dollar per wattTokens processed (trillions of tokens)2025Google 9.7 Tr1400TrExplosive Token GrowthSource:2025 AI Infra SummitAgentic AITodayToke

5、ns generated per queryComplexitySimple chatbot1xKnowledge assistant1xChain of thoughtUnique compute requirements per step10 x*Agentic systemUp to 100 x*Not actual numbers,approximate tokens per query based upon external sources and Intel internal estimatesUnique ModelUnique StepsVertically integrate

6、d networking&softwareDeployed on homogeneous systemsPoor performance per dollarEarly AI AgentOne LLMVendor-specific software&hardwareToken cost($/token)Performance(tokens/s)Acceptable performance regionBetter perf/$Incumbent solutionsHomogenous,vertically integrated Hard to scale perf/dollarExponent