A 409.6 GOPS and 204.8 GFLOPS Mixed-Precision Vector Processor System for General-Purpose Machine Learning Acceleration

As machine learning (ML) technology has become critical in various artificial intelligence (AI) applications such as image classification, object detection, and natural language processing, the latest ML models require massive computations and various operations to attain high accuracy, as shown in Fig.1. In order to cope with the computational challenges, many hardware accelerators have been proposed [1-4]. However, most of them have focused on a specific or a few models in the target domain, as shown in the table, only to achieve high performance without considering generic usages for various applications. With little programmability, the previous accelerators are often unable to adapt to model updates and algorithmic changes or suffer from low utilization if they adopt a new model. In addition, many support only fixed-point data types and arithmetic units, which limits their usage to emerging ML models.