Abstract:The smart edge nodes require efficient matrix-vector multiplications for local deep neural network (DNN) inference. Benefiting from its high density and CMOS compatibility, the eDRAM-based computing-in-memory (CIM) [1] –[4], especially with multi-level cells (MLCs) [4], attracts rising attention. However, the performance of prior MLCeDRAM CIM is severely limited by the inconsistency of weight representations during the programming and computing: weights are programmed as fixed voltages while transistor currents are used for computation. Thus, the programming of MLCs requires calibration due to the nonlinear transistor I-V, which can be extremely complicated in the presence of $V_{T H}$ variations. Furthermore, the computing precision is severely degraded by $\mathrm{V}_{\mathrm{TH}}$ variations when small computing currents are used for high parallelism. To fundamentally surmolunt this dilemma, we propose the first currentprogramming eDRAM CIM that unifies the weight programming and computing in the current domain. The enabling technique is a novel 3T1C eDRAM cell (Fig. 1, top right). It confers several key merits: 1) the cell is programmed by the weight current directly with the selfcalibrated voltage generated on the storage capacitor; it essentially stores the weight current instead of a fixed voltage, thus mitigating $V_{\text {TH}}$ variation and nonlinear transistor I-V impacts; 2) a dynamiccascoded read structure is proposed to significantly reduce the V B sensitivity while not requiring any bias voltage; 3) thanks to the accurately programmed cell, it supports MLC operation (8 current levels) without any calibration, largely increasing density; 4) a voltage-current two-step programming scheme significantly boosts the sub- $\mu \mathrm{A}$ current-weight writing speed. Combining these merits, the proposed eDRAM cell is naturally immune to transistor-level nonidealities, thus allowing a small LSB weight current of only 100nA. A $4 \mathrm{~b}$ CIM cell composed of 2MLCs is developed to support 4bsigned weights. It contains 15 current levels ranging from -700nA to 700nA. Fabricated in a $65 \mathrm{~nm}$ CMOS, the prototype achieves the highest macro-level 4b-MAC energy efficiency of 233-305TOPS/W among eDRAM CIMs.

Low-Power and Scalable Retention-Enhanced IGZO TFT Edram-Based Charge-Domain Computing

Low-Power and Scalable BEOL-Compatible IGZO TFT eDRAM-Based Charge-Domain Computing

An 8-Bit in Resistive Memory Computing Core with Regulated Passive Neuron and Bitline Weight Mapping

A Low-Power In-Memory Multiplication and Accumulation Array with Modified Radix-4 Input and Canonical Signed Digit Weights

A Robust 8-Bit Non-Volatile Computing-in-Memory Core for Low-Power Parallel MAC Operations.

High-density Energy-Efficient Charge-Domain Computing Based on CAA-IGZO TFT with BEOL-compatible 3D Integration

Low Power and Temperature-Resilient Compute-In-Memory Based on Subthreshold-FeFET

Counteractive Coupling IGZO/CNT Hybrid 2T0C DRAM Accelerating RRAM-based Computing-In-Memory Via Monolithic 3D Integration for Edge AI

Monolithic 3-D Integration of Counteractive Coupling IGZO/CNT Hybrid 2T0C DRAM and Analog RRAM-Based Computing-In-Memory

A Hybrid Computing-In-Memory Architecture by Monolithic 3D Integration of BEOL CNT/IGZO-based CFET Logic and Analog RRAM

Enabling Lower-Power Charge-Domain Nonvolatile In-Memory Computing with Ferroelectric FETs

Area-and Power-Efficient Compute-in-Memory Macro Based on Oxide Semiconductor 2T0C Edram with Cell-System Co-Optimization

A 28nm 8928Kb/mm 2 -Weight-Density Hybrid SRAM/ROM Compute-in-Memory Architecture Reducing >95% Weight Loading from DRAM.

TFT-Based Near-Sensor In-Memory Computing: Circuits and Architecture Perspectives of Large-Area Edram and ROM CiM Chips

30 Mb/mm<sup>2</sup>/layer 3D eDRAM Computing-in-Memory with Embedded BEOL Peripherals and Local Layer-wise Calibration based on First-Demonstrated Vertically-stacked CAA-IGZO 4F<sup>2</sup> 2T0C Cell

Almost-Nonvolatile IGZO-TFT-Based Near-Sensor In-Memory Computing.

A Calibration-Free 15-Level/cell Edram Computing-in-Memory Macro with 3T1C Current-Programmed Dynamic-Cascoded MLC Achieving 233-to-304-tops/w 4b MAC

Capacitor-less, Long-Retention (>400s) DRAM Cell Paving the Way towards Low-Power and High-Density Monolithic 3D DRAM

An Energy Efficient Computing-in-Memory Accelerator With 1T2R Cell and Fully Analog Processing for Edge AI Applications

Cryogenic quasi-static embedded DRAM for energy-efficient compute-in-memory applications