Font Size: a A A

Research On Compiling Implementation Of CMOS Computing-in-Memory Circuit

Posted on:2022-02-07Degree:MasterType:Thesis
Country:ChinaCandidate:Z D HuangFull Text:PDF
GTID:2518306524986959Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
Recently,neural networks has been used in various ways of daily life.Compared with the biological neuron model,the artificial neural network based on the traditional computer architecture needs more hardware resources.One of the reasons is that due to the limitation of the traditional Von Neumann architecture.In this regard,researchers put forward the concept of Computing-In-Memory(CIM),which directly completes part of computing operations in memory,thus improving data throughput and performance.The design of CIM circuit is usually built by hand,which is a time-consuming and tedious work.The structure of CIM circuit is similar to the traditional memory,which is composed of fixed modules and has a regular structure.Therefore,the memory compiler can be used to build the CIM circuit.It can quickly generate the CIM array circuit with variable size configuration,which is convenient for its research and application.This paper introduces the background of neural network research firstly.In the context of the challenge of the Von Neumann bottleneck in neural network hardware,the integrated architecture of memory and computing comes into being.Then the present theory of CIM circuit and memory compiler are introduced.The method of memory compiler can be used to generate regular circuits such as CIM array in software mode.This paper mainly carried out the following work:(1)In this paper,a CIM circuit based on SRAM and its key read-write modules are proposed.The array can be used for point multiplication in binary pulse neural network of MNIST handwritten data recognition.The principle and structure of the storage part of the CIM is the same as that of the standard SRAM circuit.When reading out,the stored value and the input value are multiplied and then read out,so as to realize the large-scale multiplication operation required by the proposed neural network algorithm.And other functional circuits for reading and writing are designed..(2)The open source memory compiler Open RAM is used to generate the physical files of the CIM circuit.Based on 55 nm technology,the circuit and layout of each module of the CIM array are designed to form the basic unit library.By modifying the compiler,it can use the designed basic cell library to generate the layout file,lef file,spice net list file and other physical files of CIM array circuit.(3)The compiler openram is used to generate Verilog,lib and other logical information files,and CIM module is used as the IP core.It is verified that the integrated memory and computing circuit file generated by the compiler is complete,and the semi customized method can be used to call the CIM array as IP and design its peripheral logic.From the algorithm background to the circuit and then to the compiler,uses feasibility of the compiler technology on the integration of CIM.Compared with the designed by hand,using compiler has the advantages of fast generation speed and automation.This is helpful to the research and application of the CIM circuit.
Keywords/Search Tags:spiking neural networks, computing-in-memory, memory compiler, physical implementation
PDF Full Text Request
Related items