使用 RISC-V 优化自定义工作负载.pdf

1、Optimizing Custom Workloads with RISC-VLudovic Henry-QConSF 2024Software Engineer&Team Lead atManaged Runtimes,System Libraries,ProfilingLanguage Runtimes WG Lead at“collaborative effort.to accelerate the development of open source software for the RISC-V architectureOpenJDK,Go,Python,.NET,Javascrip

2、t,Rust,WebAssemblyCompilers,Runtimes,and Ecosystem(libraries,tools)AboutAI is everywhere:search engines,autonomous systems,IoTAIs Expanding Influence and Hardware Demandshttps:/ourworldindata.org/grapher/exponential-growth-of-computation-in-the-training-of-notable-ai-systesAIs Expanding Influence an

3、d Hardware DemandsLimited availability and high cost of GPUsNVIDIA DGX H100 is 250k$and consume 120kW/rackMulti-year lead time to buyHard to come by on CloudProprietary options:NVIDIA,AMD,Intel,Google,AWS,MicrosoftCurrent Solutions and Their LimitsCurrent Solutions and Their LimitsDominance of GPUs

4、and challengesNVIDIA is worth 3.6T$Fast GPUs and interconnectGreat software supporthttps:/ Solutions and Their LimitsDominance of GPUs and challengesCUDAs moat is eroding but still strongCompetitors are making progressCurrent Solutions and Their LimitsDominance of GPUs and challengesCosts are high:1

5、00k GPUs is multi-Billion dollarsAnd thats only the CapExCurrent Solutions and Their LimitsNeed for adaptable,cost-efficient,and open solutionsOptimized for the needs at handMore competition,lower costs,more availabilityRISC-V:A Modular,Open ApproachOpen-Standard ISA:free to use and customizeRISC-V

6、International(RVI)4600 members:companies,universities and individualsRISC-V:A Modular,Open ApproachScales from IoT to supercomputersForecast:20B SoC shipped,to exceed 25%market share in 2031From microcontrollers to supercomputersRISC-V:A Modular,Open ApproachGrowing ecosystem with contributions to S