A high-precision stochastic computational multiplier with low input requirements
To address the problem that the output accuracy of the traditional stochastic computation multiplication was affected by the correlation between the input bit streams,a high-precision multiplier insensitive to the correlation was proposed.Firstly,the multiplication computation model was established by converting the encoding weights and combining the inputs.Secondly,a non-scaling optimization circuit was designed to solve the scaling output problem of the stochastic computation multiplier circuit and improve the computational accuracy.Finally,several distributed bit streams were used as inputs to verify the adaptation of the multiplier to the bitstreams;the effectiveness and fault tolerance of the multiplier were verified by introducing bit flipping and performing Gaussian filtering experiments.The simulation results show that the multiplier outperforms the traditional stochastic computational multiplier in terms of accuracy increasement of 54.8％and area reduction of 47.49％.In comparison to the binary multiplier,the area is saved by 82.50％.
万方数据
维普
