Comparative Analysis of Hardware Performance for Linear Detection in a Massive MIMO System on FPGA Using the Vivado HLS Tool

Abstract

This paper compares the performance of hardware implementation for linear detection in a massive MIMO system. The study focuses on Gram matrix inversion solved using two approaches: direct and indirect matrix inversion. Direct matrix inversion is represented by Cholesky Decomposition, while indirect matrix inversion is represented by the Neumann series and the Gauss-Seidel method. The algorithm for inversions, embedded in a C-based function, is virtually implemented on the FPGA using the Vivado HLS tool. The synthesis report categorizes the performance from the FPGA implementation into three parts: timing (ns), cycle latency, and resource utilization. With the same targeted time limit, indirect matrix inversion such as the Neumann series seems to be the fastest algorithm compared to the direct method due to the matrix-matrix multiplication approach. In terms of latency, NS requires more clock cycles to obtain the output compared to others.  Based on the results, the direct inversion method exhibits higher complexity, particularly in timing for clock frequency and resource utilization needed to complete the inversion