Convolution acceleration unit designing method based on chemical reaction network

Abstract

The invention discloses a convolution acceleration unit designing method based on a chemical reaction network. The method herein includes the following steps: designing N-parallel fast convolution units, also recording function modules which are used; inputting xi and hi in the form of reactant concentration sequence, designing a corresponding reactant; observing the values of logical circuit branches at different moments and corresponding reactants or products; tracking the value of a variant of respective branch; in case of a negative value of a branch, combining modules around the branch, guaranteeing the combination outputs a positive value, transplanting the combination entirely to the chemical reaction network; after tracking the entire circuit, combining all the related chemical reaction, and completing the convolution acceleration unit based on the chemical reaction network. According to the invention, the method herein simplifies and reconstructs a convolution kernel to a fast convolution unit and re-maps the fast convolution unit to the chemical reaction network, realizes high level of concurrent operation and user friendliness, and provides basis for future applications.

Description

technical field [0001] The invention relates to the field of chemical reaction network calculations, in particular to a method for designing convolution acceleration units based on chemical reaction networks. Background technique [0002] In view of the fact that Moore's Law is constantly approaching its physical limit, people have gradually turned their attention to other new computing fields, such as photonic computing, quantum computing, and molecular computing. Among them, molecular computing, especially DNA computing, has attracted much attention because of its nanometer size, strong operability, large storage capacity, and strong affinity with the human body. Chemical reaction networks (CRNs for short) are composed of a series of The set of elementary reactions (including reactants, products, and reaction rate constants) of . In molecular computing, CRNs are not only a modeling language for describing the behavior of chemical systems and constructing dynamic models,...

AI Technical Summary

Problems solved by technology

[0003] However, the convolution kernel used by the popular CNN algorithm in the neural network is a traditional silicon-based hardware circuit. The degree of parallelism is not high, the operation speed is slightly slower, and the efficiency can wait for further improvement.

Images