Prototype Verification Of Embedded Vision Processor Based On FPGA Platform

With the popularization and application of deep learning algorithms,higher requirements on computing power is put forward by artificial intelligence,since traditional CPU architecture cannot meet the demand for computing power of deep learning.Therefore,artificial intelligence chips with massive data parallel computing capability and acceleration of computing processing came into being.For the large and complex design,it has become one of the focuses of research in the IC industry at home and abroad to find a more efficient and complete verification scheme.This article proposed a system-level application verification of an embedded vision processor,which aims to help verify the correctness of the processor system functions and provide some performance data to help designers optimize and improve the design.Through the analysis of the current mainstream verification technologies,combined with the resources provided by the internship company,an efficient verification scheme is explored.Based on the HAPS-80 S26 FPGA prototype verification platform,combined with software simulation and Ze Bu debugging verification methods,the software and hardware collaborative verification of the convolutional neural network algorithm supported by the processor is performed.The software simulation verification platform uses a coverage-driven random verification method to comprehensively verify the HLAPI of CNN Engine,to make the functional verification rate of the final verification reaches 100%,which proves the correctness of the processor's function.For the FPGA prototype verification platform,after a code migration operation,a certain processor has 30 MHz the clock frequency of,and the margins of the worst setup time and the worst hold time are 0ns and 0.023 ns,respectively,which meet the timing requirements.The utilization of FPGA resources has not exceeded 50%.The total power consumption is 5.104 W,in which the static power consumption is 2.732 W,and the dynamic power consumption is 2.373 W.The mainstream 35 convolutional neural network models were tested to have correct results.It also obtain a good benchmark test result.Taking AlexNet as an example,the FPS can reach 2.21 under the test of 120 images.The verification results show that the design idea of the FPGA prototype verification platform is correct.Finally,the debugging function of the ZeBu simulation platform is explored.Three waveform dump methods,namely Dynamic-Probes,QiWC and FWC,are summarized,which effectively improve the efficiency of processor verification.This article successfully completed the verification of the expected verification goals.