Multilinear Generalized SVD Processor for Multicast Multiuser MIMO Systems

The design of the precoder and decoder is an essential technology for massive multi-user MIMO (MU-MIMO) systems because it must efficiently allocate the total channel call to all users. Generalized eigenvalue decomposition (GEVD) is a critical channel orthogonalization processing for multi-user MIMO precoding and decoding algorithms. However, the GEVD only supports unicast MU-MIMO systems. Therefore, this paper presents a multilinear generalized singular value decomposition (ML-GSVD) precoder chip, a higher-order extension of generalized singular value decomposition (GSVD), for the MU-MIMO systems that support private and shared data streams in the 5G systems. ML-GSVD can overcome the limitation of user numbers and support multiple antennas for each user by relaxing the restriction on the number of matrices, facilitating broadcast, multicast, and unicast transmissions. This paper proposes a tensor-based partitioned alternating least squares (PALS) algorithm for ML-GSVD, which significantly reduces the computational complexity of the traditional alternating least squares (ALS)-based direct-fitting approach. Through the partition- and tensor-based design strategies, the ML-GSVD computes only a portion of the matrix in each iteration and facilitates the modular and pipeline processing for hardware implementation. This study designed and implemented the proposed hardware-efficient PALS ML-GSVD algorithm using TSMC 40nm CMOS technology. The IC synthesis results show that the proposed ML-GSVD processor chip achieves a throughput of 824.27 matrices per second with a 200 MHz operating frequency.