The Design And Hardware Implementation Of MIMO Detector Based On Approximate Computation

Massive Multiple-Input Multiple-Output(MIMO)technology can more significantly improve the capacity,reliability and spectral efficiency of wireless communication sys-tems,successfully ranked as one of the core technologies of the Fifth Generation(5G)mobile communications.However,with the development of communication technology,the continuous increase in antennas poses a significant challenge for the signal detection of large-scale MIMO systems.The computational complexity of both traditional nonlinear and linear detection methods is exponentially related to the number of antennas,and the corresponding hardware implementation cost is very high,which is not realizable in prac-tical massive MIMO systems.This limits the development of massive MIMO technology to a certain extent.In this thesis,we aim to achieve a low-complexity and low-overhead massive MIMO detector to guarantee performance.Based on the approximation calculation,we apply a gradient descent iterative approximation scheme to overcome the problem of the exces-sive complexity of the massive MIMO detection algorithm and design a low-overhead hardware scheme jointly with mixed precision and dynamic stochastic computation to re-alize the joint optimization of algorithm and circuit.The innovation points of this paper are summarized as follows:(1)A low-complexity MIMO iterative detection algorithm based on gradient descent is proposed.The algorithm avoids the problem of the high complexity of directly com-puting the inverse of a matrix with high dimensionality by iterating.It takes the detection matrix as the output target to obtain the explicit expression of the detection matrix directly.Compared with existing iterative detection methods,the iterative initial value of this al-gorithm does not require a priori information and avoids additional related calculations.Simulations are performed in a scenario considering user interference in the neighbour-hood.The results show that the algorithm reduces the computational complexity from O(N_r~3)to O(mN_r)with the detection performance approximating the exact MMSE-IRC detection performance.(2)An initial value optimization strategy and a mixed precision optimization scheme are proposed.The initial value optimization strategy exploits the correlation of the channel magnitude response in frequency,and the output result on the current subcarrier is used serially as the initial value for the next iteration on the next subcarrier,providing certain prior information for the subsequent iterative process.Compared with the case without prior information,the number of iterations of this initial optimization strategy can be re-duced by about 50%.On the other hand,the two mixed-precision schemes use different data representation bit-widths in different iteration stages or different computational parts to introduce low-precision quantization with less performance loss.The overhead analysis shows that the mixed-precision scheme can reduce the hardware overhead by 20%-50%compared to the fixed-point high-precision scheme.(3)A low-complexity hardware implementation scheme based on dynamic stochas-tic computation is proposed.On the one hand,a basic low-complexity complex multi-plier is designed according to the characteristics of the used signal.On the other hand,a low-complexity computation unit for gradients is designed based on dynamic stochas-tic computation.Compared with the conventional hardware scheme,the overhead of the proposed hardware scheme is reduced from O(n~2)to O(n)for n-bit quantization.