Reconfigurable Compute-In-Memory on Field-Programmable Ferroelectric Diodes

Abstract

The deluge of sensors and data generating devices has driven a paradigm shift in modern computing from arithmeticlogic centric to data-centric processing. Data-centric processing require innovations at the device level to enable novel compute-inmemory (CIM) operations. A key challenge in the construction of CIM architectures is the conflicting trade-off between the performance and their flexibility for various essential data operations. Here, we present a transistor-free CIM architecture that permits storage, search, and neural network operations on sub-50 nm thick Aluminum Scandium Nitride ferroelectric diodes (FeDs). Our circuit designs and devices can be directly integrated on top of Silicon microprocessors in a scalable process. By leveraging the field-programmability, nonvolatility, and nonlinearity of FeDs, search operations are demonstrated with a cell footprint <0.12 μm 2 when projected onto 45 nm node technology. We further demonstrate neural network operations with 4-bit operation using FeDs. Our results highlight FeDs as candidates for efficient and multifunctional CIM platforms.

Ferroelectric Nitride Heterostructures on CMOS Compatible Molybdenum for Synaptic Memristors

“…The programmable multi-state conductance/resistance in such epitaxial GaN/ScAlN heterostructures is a result of the new domain (with opposite polarization) nucleation, growth, and volume concentration change during ferroelectric switching. , The extracted nonlinearity and corresponding asymmetry parameters for the potentiation and depression curves are α p = 2.85, α d = −6.32, and | α p – α d | = 9.17, respectively. This nonlinearity/asymmetry behavior is due to the nonlinear change of the volume concentration of new domains when applying identical voltage pulse, which can be improved by modifying the training pulse, such as stepwise pulse . Multiple voltage pulse train tests have been performed on different memristors, showing highly repeatable potentiation and depression processes on the same device as well as on different devices (Supporting Information, Figure S8).…”

“…As such, Mo has been commonly used as the bottom electrode for high frequency acoustic filters and resonators. , Therefore, epitaxially grown III-N heterostructures on Mo will provide a viable path to achieve not only CMOS compatible ferroelectric nitrides and fully nitride-based complementary circuits but also a new class of integrable, ultralow loss, and ultrahigh frequency acoustoelectronic devices. This strategy also offers a vital path for the tight integration of nonvolatile ferroelectric ScAlN memristors with processors, which is a building block for rapid data transmission and analysis in artificial neural networks. , …”

Scandium-Doped Aluminum Nitride for Acoustic Wave Resonators, Filters, and Ferroelectric Memory Applications

“…At the same time, the device shows memristive behavior with an on/off ratio of 5 × 10 4 between a low resistance state and a high resistance state and a retention time of over 1000 s without significant degradation. Using the Al 0.64 Sc 0.36 N-based MFM diode as the basic unit, nonvolatile transistor-free ternary content addressable memory (TCAM) was demonstrated as a viable building block that can support search operations in situ memory with a reduced search delay time of <0.1 ns and a cell footprint of <0.12 μm 2 . Additionally, 16 distinct conductance states can be programmed by stepwise voltage pulse modulation in the ferroelectric diode cell, enabling it to construct arrays for deep neural network inference.…”

“…Using the Al 0.64 Sc 0.36 N-based MFM diode as the basic unit, nonvolatile transistor-free ternary content addressable memory (TCAM) was demonstrated as a viable building block that can support search operations in situ memory with a reduced search delay time of <0.1 ns and a cell footprint of <0.12 μm 2 . 90 Additionally, 16 distinct conductance states can be programmed by stepwise voltage pulse modulation in the ferroelectric diode cell, enabling it to construct arrays for deep neural network inference. Beyond, in 2021, Liu et al demonstrated that the ferroelectric AlScN could be used in constructing FeFET.…”

Graphene/MoS2−xOx/graphene photomemristor with tunable non-volatile responsivities for neuromorphic vision processing

“…The human vision system has a powerful capability in visual perception only consuming less than twenty watts of power. Such features are mainly attributed to the simultaneous sensing and early processing of visual information in the retina and parallel processing in the visual cortex [1][2][3][4][5][6][7][8] . For example, to efficiently discard the redundant visual data and accelerate subsequent processing tasks in the visual cortex, the human retina can extract critical features of visual data with plastic positive and negative photoresponse 9 .…”