An Efficient Hierarchically Parallel GPU-Accelerated Framework in Large-Scale MIMO Detection
Year: 2026
Type: conference-paper
Abstract: Large-scale multiple-input multiple-output (MIMO) systems offer high spectral efficiency but introduce significant computational challenges compared to smaller MIMO systems. A key bottleneck is the high-dimensional matrix inversion used for detection tasks. Existing matrix inversion methods are often limited by insufficient computational efficiency and poor scalability, which compromise the achievable spectral efficiency gains in MIMO.In this paper, we propose ParaNeumann, a GPU-accelerated hierarchical parallel framework for near-real-time matrix inversion in large-scale MIMO detection. The core idea of ParaNeumann is to reformulate the inversion process into a three-level parallel architecture, spanning inter-subcarrier, inter-submatrix, and intra-matrix granularities to fully exploit GPU computational resources. Within this hierarchical structure, we integrate block-diagonal structures into the Neumann series to accelerate convergence, and leverage low-precision arithmetic to reduce memory cost and improve throughput. Evaluation results demonstrate that for 512 × 512 MIMO scenarios, ParaNeumann incurs a negligible performance loss of only 0.5 dB compared to exact matrix inversion at a Bit Error Rate (BER) of 10−5. In terms of computational efficiency, our proposed framework requires as little as one-third of the computation time compared to the cuBLAS-based baseline on a commercial RTX 4090 GPU.
Institution: Peng Cheng Laboratory
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224933154