CNN accelerator configuration method, system and device

Abstract

The invention discloses a CNN accelerator configuration method which comprises the steps: obtaining a data demand range of a current system mode; obtaining a data state of a CNN accelerator corresponding to the current seed data flow; judging whether the current data state meets the current data demand range or not; if not, replacing the current seed data stream to dynamically configure the PE part in the CNN accelerator until the current data state meets the current data demand range. According to the configuration method, the PE part of the CNN accelerator is dynamically configured accordingto different seed data streams, the performance of the CNN accelerator can be dynamically adjusted according to the function requirements of different system modes, the compatibility is high, and theCNN accelerator has a wide application market. Correspondingly, the invention further discloses a CNN accelerator configuration system and device with the same beneficial effects.

Description

technical field [0001] The present invention relates to the field of CNN accelerators, in particular to a CNN accelerator configuration method, system and device. Background technique [0002] Currently, the technology for implementing a CNN (Convolutional Neural Network, Convolutional Neural Network) accelerator based on an FPGA (Field Programmable Gate Array, Field Programmable Gate Array) has become mature. see figure 1 As shown, the basic composition of FPGA-based CNN accelerator includes input cache, PE (Processing Element, processing unit), output cache and external storage, where PE is the basic calculation unit of convolution, including adder and multiplier. The PE part of the CNN accelerator is a multi-layer convolutional network formed by multiple PE units, including convolutional layers and fully connected layers. [0003] In the prior art, CNN accelerators mostly use CPU (Central Processing Unit, central processing unit) or GPU (Graphics Processing Unit, graphi...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, CNN accelerators mostly use CPU (Central Processing Unit, central processing unit) or GPU (Graphics Processing Unit, graphics processing unit) to train convolutional neural networks, and then obtain optimal setting parameters, and FPGA is only used for CNN accelerators The execution inference of the CNN accelerator is fixed, and the performance is single, which cannot meet the power consumption or performance requirements in a variety of specific environments.

Images