Method, system and device for accelerating neural network calculation and storage medium

Abstract

The invention provides a method, system and device for accelerating neural network calculation and a storage medium, and the method comprises the steps: responding to the power-on of a board card, executing a starting code in a read-only memory, and detecting whether a preset address is provided with a flag bit or not; in response to a set flag bit of the preset address, downloading input data of the neural network, model parameters and parameters required by a neural network acceleration module into a DDR or an SRAM through PCIe; in response to a received calling command, reading feature parameters and weight parameters of a convolutional layer from the DDR or the SRAM, and performing neural network operation according to the feature parameters and the weight parameters; and notifying the host that the task is completed in response to the received interrupt signal indicating that the operation is completed. The RISC-V is used as a general processor, the special neural network accelerator is designed, and the RISC-V and the special neural network accelerator cooperate to accelerate the neural network.

Description

technical field [0001] The present invention relates to the field of neural networks, and more specifically, to a method, system, device and storage medium for accelerating neural network calculations. Background technique [0002] As the research on neural network algorithms continues to make significant progress, the amount of calculation required by the algorithm has also increased exponentially, posing a huge challenge to the computing power of hardware devices. In the current situation where Moore's Law is becoming more and more difficult to maintain, it is becoming more and more unfeasible to increase computing power only through the progress of semiconductor production technology. Under the current technological level, how to improve the hardware architecture is becoming more and more important for improving computing power. Hair is important. [0003] The versatility of general-purpose processors limits their efficiency when completing special tasks. At present, it ...

AI Technical Summary

Problems solved by technology

However, ASIC requires a lot of time for design and verification, making development difficult and costly

Images