A machine learning enhanced approximate message passing massive MIMO accelerator

Machine learning in the physical layer of communication systems currently receives much attention due to its potential to improve performance over difficult or unknown channels. Model-driven machine learning combines well-established algorithms with machine learning enhancements to realize these performance gains while keeping computational complexity within practical limits. In this work, we present the first model-driven machine-learning accelerator based on Orthogonal Approximate Message Passing (OAMP) for massive MIMO. The accelerator is configurable to support various machine learning enhancements such as those used in the OAMPNet and MMNet algorithms. The accelerator architecture is implemented as a deep pipeline to maximize throughput and we explore a range of antenna, user, and modulation configurations. Our results show the feasibility of deploying machine learning enhanced algorithms in future physical layer processors.