当前位置:网站首页>Co create a collaborative ecosystem of software and hardware: the "Joint submission" of graphcore IPU and Baidu PaddlePaddle appeared in mlperf

Co create a collaborative ecosystem of software and hardware: the "Joint submission" of graphcore IPU and Baidu PaddlePaddle appeared in mlperf

2022-07-04 22:38:00 Paddlepaddle

This article has been published on the official account of the flying oar , Please check the link :
Create a collaborative ecosystem of software and hardware :Graphcore IPU With Baidu PaddlePaddle “ Joint submission ” Appearance MLPerf

AI The most famous industrial benchmark list in the field MLPerf Training 2.0 On 6 month 30 Official release . In this list , Baidu PaddlePaddle and Graphcore cooperation , stay MLPerf It opens an innovative “ Joint submission ” Pattern .

“ Joint submission ” Appearance MLPerf

This time MLPerf Training v2.0 in , Baidu PaddlePaddle and Graphcore Cooperation has an innovative result submission mode :Graphcore Use the same software and hardware configuration as Baidu (Graphcore IPU And the propeller deep learning framework ) submitted MLPerf BERT Achievements of the model . chart 1 Shows this Graphcore Use with Baidu IPU The submitted BERT Benchmark results , The performance of the propeller is similar to Graphcore Highly optimized self-developed framework PopART Performance is quite .

 picture

chart 1:MLPerf Training v2.0, Graphcore Submitted with Baidu BERT Model in Graphcore IPU Training performance results on . among , Red representative Graphcore Use the native framework PopART Achievements , Blue stands for Graphcore Achievements achieved by using the propeller framework with Baidu .

“ Joint submission ”: A new attempt of the co creation plan of the propeller hardware ecosystem

Baidu PaddlePaddle in WAVE SUMMIT 2022 In depth learning developer Summit , Together with more than ten hardware manufacturers, we have released the hardware ecosystem CO creation plan . Flying OARS will combine the features of partners' own software and hardware development stack , For different application scenarios and products , Jointly launch the manufacturer's customized version of the propeller framework for the majority of developers 、 Build an open source model library 、 Develop courses and training contents, etc , Better serve developers , Achieve ecological prosperity and win-win results .

Baidu PaddlePaddle and Graphcore Cooperative MLPerf Joint submission , It is an innovative attempt of the co creation plan of the propeller hardware ecosystem . The two companies use the same software and hardware configuration to submit grades , This joint submission is in MLPerf For the first time in the competition . So , We are in the preparation stage with the organizer MLCommons Made detailed communication , To determine the feasibility of this submission . The results of this cooperation are gratifying , Not only the achievements of the technical cooperation between the two sides passed MLPerf Introduce to global developers , This soft and hard cooperation submission mode is also in MLPerf The media communication meeting was praised by representatives of other manufacturers .

“ Joint submission ” Behind the technical cooperation

For this time MLPerf Joint submission , Baidu PaddlePaddle and Graphcore Deep collaborative optimization , It is mainly reflected in the following aspects :

Optimize parallelism

Optimize the model segmentation strategy , So as to improve the parallelism of the model , And by optimizing the strategy of reading data sets in parallel , Put the model in IPU The throughput on the is brought into full play .

Improve on-chip memory utilization

By reducing the accuracy of some operations, compress the memory occupation on the chip , Save on-chip storage space , Thus, it supports the migration of optimizer state from off chip storage to on-chip storage , Reduce on-chip and off chip IO Interaction , And it can also improve the proportion of on-chip memory used by some operators , Improve the computational efficiency of operators .

The fusion collective operator

Multiple data generated in parallel collective Operators fuse into a single operator , While reducing synchronization overhead , It can also increase the utilization of bandwidth , Develop IPU Efficient computing performance .

Reduce unnecessary computing resources

Originally in evaluation Invalid reverse calculation is required in the process of , Currently, only the properties of the forward graph are calculated , save evalution Invalid calculation resource occupation in the process , Improve the whole evalution performance .

Hardware optimization

Bow-2000 Compare with M2000 It has a higher dominant frequency (1.4x), Greatly improve the computational efficiency .

Flying oars and Graphcore The course of cooperation

The propeller has been actively cooperating with hardware manufacturers to optimize user experience and performance .2020 year , The oars join hands 13 Hardware manufacturers initiated “ Hardware ecosystem ”,Graphcore That is, one of the initial members .2021 year , Baidu PaddlePaddle has achieved in Graphcore IPU Comprehensive support for training and reasoning , And open source the relevant code . Both teams are in IPU-POD16&64 Data parallel and model parallel on , And in Bert-Base Intensive reading and throughput verification are carried out on the model, and good performance results are obtained .Graphcore Of Poplar SDK 2.3 With the latest Baidu PaddlePaddle framework 2.3 Version has been fully integrated , The relevant code has been in Baidu PaddlePaddle GitHub Online for developers to get .

2022 year 5 month ,Graphcore stay WAVE SUMMIT 2022 At the deep learning developer summit, it was officially announced to join the hardware ecosystem CO creation plan initiated by Baidu PaddlePaddle .Graphcore And Baidu PaddlePaddle will jointly develop technical solutions based on the co creation plan , Collaborative customization of propeller frame , Build model base and scenario base , With “IPU+ Flying propeller ” Empower industry , Drive industry AI Transformation and upgrading .

Conclusion

With the wide application and rapid development of artificial intelligence technology in various industries , The industry has entered the stage of algorithm and hardware collaborative innovation from their independent hardware computing power drive and algorithm innovation drive . This time MLPerf Joint submission of , It is the co creation partner of Baidu PaddlePaddle and hardware ecology Graphcore Innovative attempts of cooperation . future , The propeller will create a plan through hardware ecology , With more hardware vendors , Accelerate the application of artificial intelligence , Push AI The realization process of industrial mass production .

MLPerf Introduce

MLPerf By AI Benchmark in the field of artificial intelligence initiated by world-renowned academic researchers and industry experts .MLPerf It aims to provide a fair 、 A practical benchmark platform , Show industry-leading AI The best performance of software and hardware system , The test results have been obtained AI General recognition in the field . Almost all mainstream hardware manufacturers and software service providers in the world will refer to MLPerf Build your own benchmark system based on the published results , To test the new AI Accelerating chip and deep learning framework in MLPerf Performance on the model .

Read more

Live broadcast announcement

7 month 6 Japan ( Wednesday )20:00, Yu dianhai, the chief architect of flying oars, and Zeng Jinle, the senior R & D Engineer of flying oars, will broadcast live , Uncover secrets for everyone GPU Under configuration , Baidu PaddlePaddle performance 「 World first 」 The key technology behind it .

Scan the qr code below , The background to reply 【 Study 】 Sign up , There are more gifts waiting for you in the live studio !

 picture

Focus on 【 Flying propeller PaddlePaddle】 official account
Get more technical content ~

原网站

版权声明
本文为[Paddlepaddle]所创,转载请带上原文链接,感谢
https://yzsam.com/2022/185/202207042214501488.html