面向新型计算架构的存算融合晶体管器件研究

Research on storage-computing fusion transistors for novel computing architectures

基于金属-氧化物-半导体场效应晶体管(metal-oxide-semiconductor field-effect transistor,MOSFET)器件和冯·诺伊曼架构的半导体芯片技术带领人类迈进了信息化时代.近年来,新型计算架构蓬勃发展.然而,专用算法和类脑智能算法映射至MOSFET电路时,通常面临硬件开销大、系统能效低等挑战.针对上述问题,本文研制了具有存算融合特点的新型晶体管器件,有效适配了专用算法和类脑计算应用:在基础数值计算方面,针对多项式计算,研制了电荷捕获型晶体管,利用器件的本征非线性动力学特征,基于单器件实现了三元素乘法运算,有效加速了多项式回归任务;在新兴智能计算方面,针对神经网络计算,研制了离子栅型神经形态晶体管,利用双栅耦合的方式实现了神经元的非线性激活和时空信息整合功能,并基于此提出了神经元相关性脉冲神经网络,实现了注意力转变现象的模拟.本文的工作将为新型计算架构芯片的构建提供新的思路.

With the rapid development of information technology,various innovative computing architectures are leading the evolution of semiconductor chip technology.Due to the separation of memory and computation in traditional von Neumann architecture,the process of data transfer between memory and processor has given rise to a critical bottleneck that limits computing power and energy efficiency improvement in modern chips.Consequently,the development of high-efficient chip technology with storage-computing fusion architecture has drawn great attention from both academia and industry.In recent years,dedicated algorithmic acceleration chips based on complementary metal oxide semiconductor(CMOS)have experienced rapid growth,and playing an important role in basic numerical computing and emerging intelligent computing fields.One noteworthy task in basic numerical computing is polynomial regression calculation,which holds importance in information encryption,image processing and various other fields.Polynomial regression calculation entails a large number of multiplication operations which need massive multiplier resources in the field programmable gate array(FPGA)or graphics processing unit(GPU),resulting in huge hardware overhead and power consumption.To settle this issue,this paper develops a charge capture transistor based on indium gallium zinc oxide(InGaZnO,IGZO)channel.Leveraging the nonlinear characteristics of the device,the ternary multiplication operation is achieved in the analog domain by a single device.By applying the proposed transistor array,the sensitive parameter analysis of liver disease problem(a typical polynomial regression task)is effectively completed.In the emerging intelligent computing field,brain-inspired neural networks have garnered substantial attention for their ability to perform complex computation with low power consumption and reduced hardware cost.As one of the core components of brain-inspired neural networks,neurons possess the capabilities of nonlinear activation computing and spatiotemporal information integration computing.However,traditional CMOS technology struggles to efficiently adapt to brain-inspired neural networks due to their singular functionality and high hardware demands.To settle this issue,this paper develops an ion-gated neuromorphic transistor.Leveraging the synergistic regulation of top and back gates,the spatiotemporal information integration characteristic of neurons is achieved by a single device.Moreover,this paper proposes a correlation-based spiking neural network based on the spatiotemporal information integration characteristic of the transistor.The proposed design carries out the simulation of attention-shifting task,providing guidelines for the construction of novel brain-inspired neural network.