Design of an SVD Engine for 8x8 MIMO Precoding Systems

A singular-value-decomposition (SVD) engine for 8 x 8 MIMO precoding systems is designed and implemented. The memory-based architecture is adopted with eight processing elements, each having two CORDIC modules. Two-phase operations are performed including bidiagonalization and Golub-Reinsch SVD (GR-SVD) with Rayleigh quotient shift. The split, deflation, and shift techniques of GR-SVD can effectively decrease the processed matrix size and accelerate the diagonalization to enhance the throughput. To cover the wide distribution of singular vector elements and singular values derived from 8 x 8 MIMO channel matrix, hybrid datapath representations are used. The thresholds for split and deflation can be adjusted and thus the accuracy of the SVD engine is variable according to the requirements. From the synthesis results, the SVD engine in 45 nm CMOS technology is able to provide the throughput rate of 636 K matrix/s and outperforms the previous design.