On-Chip Memory Optimization of High Efficiency Accelerator for Deep Convolutional Neural Networks
Tzu-Yi Lai, Kuan-Hung Chen · 2018
Artificial intelligence (AI) machine often used Deep Convolutional Neural Networks (DCNN). In this paper, we delved the relevance between memory size and convolutional network architectures, i.e., AlexNet [1] and YOLOv2[2]. The high efficiency accelerator for AlexNet uses 134 k Byte memory size. Comparably, YOLOv2 costs 127 k Byte. The type of network architecture must be analyzed to determine how to save space. From this study, we get two design tips, i.e., we should focus on the use of level-2 filter memory when realizing the DCNN models having larger filter size. Meanwhile, when implementing the models having smaller filter size and large image size, we should focus on the use of level-3 image memory and its disassembly.