Design And Implementation Of Ai Processor Acceleration Module Based On RISC-V

In recent years,artificial intelligence has been applied to more and more embedded environment,with the fact that many artificial intelligence application fields have higher requirements on performance and power consumption,usually requiring a high performance or high energy efficiency solution,traditional kinds of cpus are difficult to meet these requirements,so domain-specific processing architectures that can accelerate certain kinds of algorithms are born.However,although there are many AI algorithm hardware acceleration schemes at present,they are usually based on high power processor cores,which incur huge resource overhead and high energy consumption.RISC-V,with its features of simplicity and open source,is suitable for embedded low-power scenarios.Therefore,this thesis takes low power artificial intelligence processor as the research target,designs and implements RISC-V based artificial intelligence processor acceleration module.In order to enable the AI processor to speed up the AI algorithm,this thesis studies the RISC-V instruction expansion rules,proposes a custom AI instruction set,and realizes the support of convolution,maxpooling,activation,vector addition,and fully connection operation.The convolutional operation unit designed in this thesis uses the systolic array method and improves the traditional systolic array convolutional acceleration architecture to further explore the data-level parallelism.In this thesis,the convolutional computing unit is reused and the system resource cost is reduced.Fully connected cells are implemented using pipeline-based parallel addition tree.In addition,the auxiliary buffer module based on ping-pong buffer is designed to reduce load and store times.Finally,this thesis carries on the functional test,the acceleration ratio test,the energy efficiency ratio test to the system.The test results show that the artifical intelligence processor acceleration module based on RISC-V designed and implemented in this thesis can correctly run the AI algorithm,the acceleration ratio is qualified,the power consumption of the system is 0.41 W,energy efficiency ratio is 42.4,meet the design target.