A test case for the evaluation of AI-oriented hardware accelerators reliability

This article focuses on AI-Oriented Hardware Accelerators Reliability Using NVDLA Deep Learning Accelerator from NVIDIA As an example, the NVDLA Deep Learning Accelerator was designed using zoix and terata max software. There are six main parts in the essay, namely introduction, background, implementation, result, discussion and conclution and appendix. introduction introduces the current problem to be solved, a short summary and the structure of the essay. The background section introduces the composition of Convolutional Neural Network, the hardware composition of NVDLA, the software used in the thesis, and the methodology used in the thesis. The implementation section describes how to generate the corresponding test vectors for integer and floating-point multipliers and the corresponding addition trees. The results section shows the simulation results for the final test vector. The discussion and conclution section summarizes the focus of the thesis and its shortcomings The appendix records the scripts and code used and the test vectors generated. The following are the simulation results of generating test vectors using the method in the paper. The comparison shows that as a field test, the vectors tested using this method can basically test all possible scenarios, and the number of test vectors is smaller, which takes up less storage volume.