软件优先的方法 ai在边缘.pdf

1、Tim MamtoraChief of Innovation&Engineering,ImaginationSpeaker:A software-first approach to Huge societal opportunity for AI to“do good”Source:IDC,#US50888824,May 2024Worldwide Edge Endpoint AI Processor and Accelerator Forecast,20242028Challenges at the edge are all too familiarThe biggest lesson th

2、at can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective,and by a large margin”Two key lessons:Scalable and OpenDirectly competing with open source is a losing propositionOpen-source models are faster,more customizable,more priva

3、te and pound-for-pound more capable”Hotspot management challengesSilicon integration challenges with increasing transistor count per deviceIP selection and supply challenges beyond 2nmHeterogenous programming challengesSimpler thermal managementMore tractable silicon integration whether monolithic,2

4、D or 3DSimpler IP management scenarioIt scales!Its simple to use!Traditional Edge Device Model Domain Specific Accelerators(DSAs)and rich IP ecosystemFuture Edge Device Model Borrowing from cloud design philosophyTaking a hardware-first viewTaking a software-first viewNon-proprietary entry pointsAvo

5、idance of premature optimisationAvoidance of premature optimisationGood quality and extensive libraries and tools User can exert fine control over the hardwareSufficient degrees of freedom to address workload specific needsComputational Generality and Ecosystem CompatibilityEase-of-use and Good SW a

6、nd LibrariesSolution-level Energy EfficiencyEnter the Unified Acceleration Foundation(UXL)Build a multi-architecture,multi-vendor software ecosystem for all acceleratorsUnify the heterogeneous compute ecosystem around open standardsBuild on and expand open-sou