012-Rocky-Zhang.pdf

1、2024 SiFiveEnabling energy efficient DatacentersRocky ZhangPrincipal Engineer,Field Application,SiFiveRISC-V SummitHangzhou,ChinaAugust 21,20242024 SiFiveSummary3Datacenter workloads are diversifying and an opportunity exists to create optimized-for-purpose solutionsFeaturing a combination of genera

2、l-purpose and application-specific processing elementsThe P870-D processor represents SiFives first processor that addresses mainstream datacenter and infrastructure use caseComplementary to SiFives Intelligence Processor familyHarnessing open-source software to provide foundational platform that ad

3、dresses emerging class of AI workloads need for cost-and power-efficiencyDelivers size/power advantages over incumbent approaches Most appropriate for workloads that benefit from high levels of parallelism2024 SiFivePrimary Factors Driving RISC-V Into Higher Value Sockets4Increased use of open-sourc

4、e APIs and foundational software across use casesSustainability,cost and/or performance concerns driving need for innovation,with no one size fits allA desire from industry for increased resilience to supply chain disruption2024 SiFivePrimary Factors Driving RISC-V into Higher Value Sockets5Web serv

5、icesMedia streamingStorageData analyticsAI(where main training functionality is offloaded to GPGPU and/or hardware accelerator)High Performance Computing(HPC)Presentation LayerAnalytics and BIProcessing LayerData LayerData IngestionSensors|Logs|Social|GeoWeather|IoT|UnstructuredData Flow2024 SiFiveK

6、ey Infrastructure Requirements6TCO-High performance with great compute density/power efficiencyScalability-Creating high core count systems,across multiple die/chipletsAI Workloads-Effective coupling of optimized-for-purpose acceleratorsInfrastructure-specific features-RAS,Security etcTime to Market