Research and Development of Weight Compression Based on NVDLA

With the increasing demand for computing power by deep neural networks, the demand for memory and network resources is also growing. The majority of weight data exhibit sparse characteristics. Therefore, as a dedicated hardware accelerator for deep neural networks, the weight compression function is gradually becoming an important development trend for deep neural networks. Nvidia Deep Learning Neural Network Accelerator (NVDLA) is an open-source neural network hardware accelerator from Nvidia, which is highly recognized by academia and industry for its outstanding performance. This paper mainly studies the weight compression problem of sparse weights in neural network models for NVDLA, and through various optimization solutions, it efficiently utilizes the internal resources of NVDLA while reducing the demand for bandwidth. Based on the built NVDLA accelerator platform, this paper implements weight compression and hardware acceleration for LeNet and ResNet18 neural network models, and completes image classification tasks on the handwritten digits and CIFAR10 datasets. The research results show that NVDLA with weight compression function can significantly reduce bandwidth, with the best compression ratio up to 3.31x, while ensuring processing performance up to 36.6 fps.