High level synthesis strategies for ultra fast and low latency matrix inversion implementation for massive MIMO processing

Abstract

In this paper, we discuss the implementation strategies of an explicit matrix inversion technique based on the recursive Gram matrix inversion update (RGMIU) algorithm. These strategies are explicitly chosen to optimize the design for high throughput dictated by the enhanced mobile broad band (eMBB) use case and by the low latency imposed by the ultra-reliable low-latency communications (URLLCs) use case. The RGMIU algorithm is recently proposed to implement the zero forcing (ZF) problem encountered in massive multiple-input multiple-output (MIMO) detection task. We therefore compare and analyse the performance in terms of symbol error rate (SER) against popular implicit and explicit methods such as optimized coordinate descent (OCD), Gauss-Seidel (GS) and Neumann series expansion (NSE) algorithms. To determine the optimal word length, a fixed-point analysis shows that 16 bits are enough to reach a floating-point precision. We, thereafter, use Vivado High-Level Synthesis tools and optimization directives to implement the RGMIU algorithm based on three strategies to infer key insights and design trade-offs from the resource's utilization, latency, throughput and energy efficiency.