• 基于3T0C阵列和级联电流镜的高线性度存内计算方案设计

    Design of a high-linearity computing in-memory scheme based on 3T0C array and cascade current mirror

    • 存内计算(Compute in Memory,CIM)能够减少数据搬运,从而有效缓解冯·诺依曼架构中的“存储墙”瓶颈。其中,基于非晶氧化物半导体晶体管(Amorphous Oxide Semiconductor Field-effect Transistors,AOSFET)的嵌入式动态随机存取存储器(embedded Dynamic Random-access Memory, eDRAM)因其高集成度优势,适用于构建大规模模拟CIM阵列。相较于传统双晶体管无电容(2T0C)单元方案,三晶体管(3T0C)单元通过深度抑制交叉阵列中的潜行电流并切断读字线与位线间耦合电容形成的反馈路径,显著提升了CIM结果的线性响应度。然而,受限于读位线电容,3T0C方案在多行并行读取或高工作温度等大读电流工况下,仍面临线性响应度严重下降的问题。为此,本文提出在3T0C阵列的读位线上增设级联电流镜(Cascaded Current Mirror, CCM)输出级,以每列4个晶体管的低硬件开销,通过钳位读位线电压实现读电流与电压摆幅的解耦,并利用镜像比调控等效负载电容的放大倍数。256×64阵列上的仿真结果表明,在室温和80°C条件下,所提设计在四行并行读取时的积分非线性(INL)均小于1个最低有效位(LSB),较无CCM的3T0C方案分别改善了70%和83.8%。

       

      Abstract: Compute-in-Memory (CIM) technology can reduce data movement, alleviating the“memory wall” bottleneck inherent to the von Neumann architecture. Among various implementations, embedded Dynamic Random-access Memory (eDRAM) based on Amorphous Oxide Semiconductor Field-effect Transistors (AOSFETs) is particularly attractive for constructing large-scale analog CIM arrays due to its high integration density. Compared with the conventional two-transistor zero-capacitor (2T0C) cell, the three-transistor zero-capacitor (3T0C) cell significantly improves the linearity of CIM outputs by strongly suppressing sneak currents in crossbar arrays and eliminating the feedback path formed by the coupling capacitance between the read word line and the bit line. However, limited by the read bit-line capacitance, the 3T0C scheme still suffers from severe linearity degradation under high read-current conditions, such as multi-row parallel readout or elevated operating temperatures. To address this issue, this paper proposes a Cascaded Current Mirror (CCM) output stage added to the read bit line of the 3T0C array. With a low hardware overhead of only four transistors per column, the proposed design decouples the read current from the voltage swing by clamping the read bit-line voltage, while the mirror ratio is exploited to regulate and amplify the equivalent load capacitance. Simulation results on a 256×64 array show that, at room temperature and 80°C, the proposed design achieves an integral nonlinearity (INL) of less than one least significant bit (LSB) during four-row parallel readout, improving the INL by 70% and 83.8%, respectively, compared to the 3T0C scheme without the CCM.

       

    /

    返回文章
    返回