The present disclosure provides a calculating device for a sparsely connected artificial neural network, comprising: a mapping unit for converting input data into a storage mode in which input neurons correspond one-by-one to weight values; a storage device used for storing data and instructions; an operation unit for executing a corresponding operation on said data according to the instructions, wherein said operation unit mainly performs a three-step operation: step 1, multiply the input neurons and the weight values; step 2, execute an adder tree operation for performing, by means of an adder tree, a stage-by-stage summation of the weighted output neurons processed by step 1 or add a bias to the output neurons to obtain biased output neurons; and step 3, execute an activation function operation, to obtain final output neurons. The device resolves the problem that operation performance of a CPU and a GPU is insufficient, and an overhead of front end coding is large, effectively improving support of a multi-layer artificial neural network operation algorithm, and avoiding the problem that the memory bandwidth becomes a bottleneck for a multi-layer artificial neural network operation and the training algorithm performance thereof