In this work, we realized a three-dimensional (3D) reservoir computing (RC) by utilizing the I-V nonlinearity and short-term memory of the dynamic memristor in 4-layer vertical array. The cycle-to-cycle variation of the dynamic reservoir is improved by parallel memristor configuration. The dimensionality of the reservoir space is increased by input strategy design. After the hardware-software co-optimization, the proposed 3D RC system exhibits high recognition accuracy (90%), low energy consumption (~0.78 pJ /operation), and high area efficiency (5.12 TOPS/mm 2 ).