Characterizing convolutional neural network workloads on a detailed GPU simulator

Recent frameworks on convolutional neural networks (CNNs) such as Caffe and MXNet have focused primarily on being compatible with CUDA software and hardware application. However, it was designed for GPU architecture of compute capability 3.0 and above. Therefore, it needs verification of function to perform GPGPU-Sim which is implemented as NVIDIA compute capability devices 2.x. We developed a framework which can make inferencing AlexNet on GPGPU-Sim. We also analyze the execution results of the GPGPU-Sim. The number of lines in one set of the L1 data cache is sensitive to influence performance of AlexNet inference.