Intelligent Chip Performance Evaluation And Optimization Based On DNN

With the rapid development of the artificial intelligence technology,the big data,deep learning algorithms and hardware are also being updated every day.In terms of algorithms,various new neural network algorithms have been proposed.Most of these algorithms require higher quality data requirements.The ability of hardware is often the bottleneck of such tasks with large amounts of data and high computational.The increasing attention to deep learning has greatly stimulated the design of intelligent chips.Major companies have proposed corresponding hardware design solutions.With more and more chips,how to evaluate these chip's performances fairly and systematically and optimize these hardwares has become a research hotspot.Therefore,this paper proposes a benchmark test program for intelligent chip evaluation and analysis after evaluation to guide the optimization of hardware and software.This article first summarizes the current benchmarking program,analyzes the deficiencies,points out the existing problems,then studies and analyzes the key elements of the benchmarking program,and adds analysis methods and optimization methods to provide developers with optimization suggestions.This article mainly designs the benchmarking scheme from four aspects.First,benchmark suit design.The suite is designed from three dimensions,including microbenchmarks,mesobenchmarks,and macrobenchmarks,which are used to test hardware-based performance,analyze hardware bottlenecks,and compare hardware performance.Second,metrics design.Hardware performance is evaluated from multiple aspects,including hardware-based performance,power consumption,area,and accuracy.Third,analysis methods.Analyze from two aspects of basic testing and performance analysis,explore the hardware performance boundary based on the test results,and analyze the bottlenecks that limit performance.Fourth,the optimization method.Based on the bottleneck limitation in the analysis method,the search for optimization points is not only optimized from the hardware side,but also combined with software algorithms,and hardware and software collaborative optimization from the aspect of improving hardware utilization efficiency.Finally,this article builds a GPU verification platform based on a benchmark test scheme and verifies of the four aspects of the design of the benchmark test program.We use the benchmark suites and metrics in the benchmark test program to obtain hardware information,analyze the bottleneck of the hardware and use the optimization methods in the solution to optimize the hardware basic performance and the combination of hardware and software.