An Ultrafast (< 200 ns) Sparse Solution Solver made by HfWOx/VOy Threshold Tunable Neurons

For the first time, we develop a threshold tunable neural component based on HfWOx/VOy compound layer device and realize an ultrafast sparse solution solver with it. By mapping the activation function in locally competitive algorithm (LCA) to neural components, we construct a loop circuit capable of solving regularized linear regression problems with one-step convergence. In this novel device based architecture: (1) The ultrafast device switching speed (&lt; 30 ns) provides the key enabler for the fast convergence of the solver. (2) the solving time remains below 200 ns for problems with various scales, which is about 10(4) times faster than CMOS ASIC scheme. (3) More than 5 thresholds (V-th) can be tuned out in the HfWOx/VOy neural component with excellent endurance of 10(10) switches. Such adjustable V-th provides a highly flexible and efficient solution for application scenarios with different requirements on sparsity-accuracy balance.