MN-Core 2：面向人工智能和通用HPC应用的第二代MN-Core架构处理器.pdf

1、MN-Core 2:Second-generation processor of MN-Core architecture for AI and general-purpose HPC applicationJun MakinoPreferred Networks,Inc./Kobe UniversitySummary2MN-Core is a series of AI-oriented accelerator developed by PFNThe first-generation MN-Core has achieved very high performance-per-watt,ran

2、ked#1 in Green500 list 3 times.We completed MN-Core 2.Achieved higher performance per watt and also higher performance per silicon area.Performance on real applications,both AI and general HPC areas,is higher than that of other accelerator made with same process tech.We have already started the deve

3、lopment of next-generation processors,both for learning and LLM inference.Talk outline3Who are We?MN-Core MN-Core 2MN-Core 2 achieved performanceArchitecture of MN-Core 2Software for MN-Core 2Future directionSummaryPreferred Networks Inc.Industry DomainsTransportation ManufacturingLife SciencesMater

4、ialsRobotsEntertainmentFoundedMarch 2014DirectorsCEOToru NishikawaCERDaisuke OkanoharaLocatedTokyo,Japan(HQ)Burlingame,CA.,US(Preferred Networks America,Inc.)Vision:Make the real world computable4MN-Core:Accelerator for Efficient Deep LearningOur roadmapLaunched the project at 2017,we operate MN-3 s

5、upercomputer with the first generation accelerator MN-Core from 2020Second-generation MN-Core 2 is up and running.Future generations under developmentDesigned for efficientdeep-learning trainingand backpropagationLarge-scale SIMD processorPeak Performance:32.8TFlops(double-precision)Chip Power:500W

6、Computing Efficiency:65GFlops/WUnique ArchitectureNo data cacheOn-chip network wit tree topology and support for broadcast and reduction in all levels of the treeDesigned for tensor-oriented operations of Deep LearningThird-genForth-gen5MN-Core forLLM InferenceOur MN-3 supercomputer is certified as