UPE: A Device-Edge DNN Inference Artificial Intelligence Processor with Supporting Reconfigurable Training

For a long time, DNN inference has played a huge role in many device-edge scenarios. However, due to a lack of training capabilities, it often needs to be trained and downloaded in the cloud, which brings limitations to device-edge AI. This article proposes a device-edge DNN inference artificial intelligence processor termed UPE (Union Processing Element) which supports reconfigurable training offline and has an improved solution with three characteristics. Firstly, UPE has a DNN Training Inference Reconfigurable Computing (TIRC) architecture, which can reconstruct the computing unit from inference to training in real time to complete partial weight updates. Secondly, UPE has a Dynamic Pruning Optimization Method (DPOM), which supports separately branch pruning for different levels of inference and training. Thirdly, UPE includes the Flexible Precision Simplified Computing (FPSC) scheme, which can compress data where DNN training and numerical format inference coexist. Using 28nm CMOS process for evaluation, UPE performs well in terms of energy consumption in inference AI processor, which is also able to accomplish the training and weights offline self-updating in device-edge.