Kumar Virwani

Experimental Demonstration and Tolerancing of a Large-Scale Neural Network (165 000 Synapses) Using Phase-Change Memory as the Synaptic Weight Element

Using 2 phase-change memory (PCM) devices per synapse, a 3-layer perceptron network with 164,885 synapses is trained on a subset (5000 examples) of the MNIST database of handwritten digits using a backpropagation variant suitable for NVM+selector crossbar arrays, obtaining a training (generalization) accuracy of 82.2% (82.9%). Using a neural network (NN) simulator matched to the experimental demonstrator, extensive tolerancing is performed with respect to NVM variability, yield, and the stochasticity, linearity and asymmetry of NVM-conductance response.

Recyclable, strong thermosets and organogels via paraformaldehyde condensation with diamines

Recyclable Thermoset Polymers The high mechanical strength and durability of thermoset polymers are exploited in applications such as composite materials, where they form the matrix surrounding carbon fibers. The thermally driven polymerization reaction is usually irreversible, so it is difficult to recycle the constituent monomers and to remove and repair portions of a composite part. García et al. (p. 732; see the Perspective by Long) now describe a family of polymers formed by condensation of paraformaldehyde with bisanilines that could form hard thermoset polymers or, when more oxygenated, produce self-healing gels. Strong acid digestion allowed recovery of the bisaniline monomers. A strong polymer formed by heating can be digested with strong acid to recover and recycle its bisaniline monomers. [Also see Perspective by Long] Nitrogen-based thermoset polymers have many industrial applications (for example, in composites), but are difficult to recycle or rework. We report a simple one-pot, low-temperature polycondensation between paraformaldehyde and 4,4ʹ-oxydianiline (ODA) that forms hemiaminal dynamic covalent networks (HDCNs), which can further cyclize at high temperatures, producing poly(hexahydrotriazine)s (PHTs). Both materials are strong thermosetting polymers, and the PHTs exhibited very high Young’s moduli (up to ~14.0 gigapascals and up to 20 gigapascals when reinforced with surface-treated carbon nanotubes), excellent solvent resistance, and resistance to environmental stress cracking. However, both HDCNs and PHTs could be digested at low pH (<2) to recover the bisaniline monomers. By simply using different diamine monomers, the HDCN- and PHT-forming reactions afford extremely versatile materials platforms. For example, when poly(ethylene glycol) (PEG) diamine monomers were used to form HDCNs, elastic organogels formed that exhibited self-healing properties.

Access devices for 3D crosspoint memorya)

The emergence of new nonvolatile memory (NVM) technologies—such as phase change memory, resistive, and spin-torque-transfer magnetic RAM—has been motivated by exciting applications such as storage class memory, embedded nonvolatile memory, enhanced solid-state disks, and neuromorphic computing. Many of these applications call for such NVM devices to be packed densely in vast “crosspoint” arrays offering many gigabytes if not terabytes of solid-state storage. In such arrays, access to any small subset of the array for accurate reading or low-power writing requires a strong nonlinearity in the IV characteristics, so that the currents passing through the selecteddevices greatly exceed the residual leakage through the nonselecteddevices. This nonlinearity can either be included explicitly, by adding a discrete access device at each crosspoint, or implicitly with an NVM device which also exhibits a highly nonlinear IV characteristic. This article reviews progress made toward implementing such access device functionality, focusing on the need to stack such crosspoint arrays vertically above the surface of a silicon wafer for increased effective areal density. The authors start with a brief overview of circuit-level considerations for crosspoint memory arrays, and discuss the role of the access device in minimizing leakage through the many nonselected cells, while delivering the right voltages and currents to the selected cell. The authors then summarize the criteria that an access device must fulfill in order to enable crosspoint memory. The authors review current research on various discrete access device options, ranging from conventional silicon-based semiconductor devices, to oxide semiconductors, threshold switch devices, oxide tunnel barriers, and devices based on mixed-ionic-electronic-conduction. Finally, the authors discuss various approaches for self-selected nonvolatile memories based on Resistive RAM.

Self-assembled ferrimagnet--polymer composites for magnetic recording media.

A self-assembled magnetic recording medium was created using colloidal ferrimagnetic building blocks. Monodisperse cobalt ferrite nanoparticles (CoFe(2)O(4)) were synthesized using solution-based methods and then stabilized in solution using the amphiphilic diblock copolymer, poly(acrylic acid)-b-poly(styrene) (PAA-PS). The acid groups of the acrylate block bound the polymer to the nanoparticle surface via multivalent interactions, while the styrene block afforded the magnetic nanoparticle--polymer complex solubility in organic solvents. Moreover, the diblock copolymer improved the colloidal stability of the ferrimagnetic CoFe(2)O(4) nanoparticles by reducing the strong interparticle magnetic interactions, which typically caused the ferrimagnetic nanoparticles to irreversibly aggregate. The nanoparticle--polymer complex was spin-coated onto a silicon substrate to afford self-organized thin film arrays, with the interparticle spacing determined by the molecular weight of the diblock copolymer. The thin film composite was also exposed to an external magnetic field while simultaneously heated above the glass transition temperature of poly(styrene) to allow the nanoparticles to physically rotate to align their easy axes with the direction of the magnetic field. In order to demonstrate that this self-assembled ferrimagnet--polymer composite was suitable as a magnetic recording media, read/write cycles were demonstrated using a contact magnetic tester. This work provides a simple route to synthesizing stabilized ferrimagnetic nanocrystals that are suitable for developing magnetic recording media.

Highly-scalable novel access device based on Mixed Ionic Electronic conduction (MIEC) materials for high density phase change memory (PCM) arrays

Phase change memory (PCM) could potentially achieve high density with large, 3Dstacked crosspoint arrays, but not without a BEOL-friendly access device (AD) that can provide high current densities and large ON/OFF ratios. We demonstrate a novel AD based on Cu-ion motion in novel Cu-containing Mixed Ionic Electronic Conduction (MIEC) materials[1, 2]. Experimental results on various device structures show that these ADs provide the ultra-high current densities needed for PCM, exhibit high ON/OFF ratios with excellent uniformity, are highly scalable, and are compatible with <400°C Back-End-Of-the-Line (BEOL) fabrication.

Phase transitions in Ge-Sb phase change materials

Thin films of the phase change material Ge–Sb with Ge concentrations between 7.3 and 81.1 at. % were deposited by cosputtering from elemental targets. Their crystallization behavior was studied using time-resolved x-ray diffraction, Auger electron spectroscopy, differential scanning calorimetry, x-ray reflectivity, profilometry, optical reflectivity, and resistivity versus temperature measurements. It was found that the crystallization temperature increases with Ge content. Calculations of the glass transition temperature (which is a lower limit for the crystallization temperature Tx) also show an increase with Ge concentration closely tracking the measured values of Tx. For low Ge content samples, Sb x-ray diffraction peaks occurred during a heating ramp at lower temperature than Ge diffraction peaks. The appearance of Ge peaks is related to Ge precipitation and agglomeration. For Ge concentrations of 59.3 at. % and higher, Sb and Ge peaks occurred at the same temperature. Upon crystallization, film mass...

Experimental demonstration and tolerancing of a large-scale neural network (165,000 synapses), using phase-change memory as the synaptic weight element

Using two phase-change memory devices per synapse, a three-layer perceptron network with 164 885 synapses is trained on a subset (5000 examples) of the MNIST database of handwritten digits using a backpropagation variant suitable for nonvolatile memory (NVM) + selector crossbar arrays, obtaining a training (generalization) accuracy of 82.2% (82.9%). Using a neural network simulator matched to the experimental demonstrator, extensive tolerancing is performed with respect to NVM variability, yield, and the stochasticity, linearity, and asymmetry of the NVM-conductance response. We show that a bidirectional NVM with a symmetric, linear conductance response of high dynamic range is capable of delivering the same high classification accuracies on this problem as a conventional, software-based implementation of this same network.

Neuromorphic Computing Using Non-Volatile Memory

Abstract Dense crossbar arrays of non-volatile memory (NVM) devices represent one possible path for implementing massively-parallel and highly energy-efficient neuromorphic computing systems. We first review recent advances in the application of NVM devices to three computing paradigms: spiking neural networks (SNNs), deep neural networks (DNNs), and ‘Memcomputing’. In SNNs, NVM synaptic connections are updated by a local learning rule such as spike-timing-dependent-plasticity, a computational approach directly inspired by biology. For DNNs, NVM arrays can represent matrices of synaptic weights, implementing the matrix–vector multiplication needed for algorithms such as backpropagation in an analog yet massively-parallel fashion. This approach could provide significant improvements in power and speed compared to GPU-based DNN training, for applications of commercial significance. We then survey recent research in which different types of NVM devices – including phase change memory, conductive-bridging RAM, filamentary and non-filamentary RRAM, and other NVMs – have been proposed, either as a synapse or as a neuron, for use within a neuromorphic computing application. The relevant virtues and limitations of these devices are assessed, in terms of properties such as conductance dynamic range, (non)linearity and (a)symmetry of conductance response, retention, endurance, required switching power, and device variability. Graphical Abstract

Observation and modeling of polycrystalline grain formation in Ge2Sb2Te5

The relationship between the polycrystalline nature of phase change materials (such as Ge2Sb2Te5) and the intermediate resistance states of phase change memory (PCM) devices has not been widely studied. A full understanding of such states will require knowledge of how polycrystalline grains form, how they interact with each other at various temperatures, and how the differing electrical (and thermal) characteristics within the grains and at their boundaries combine through percolation to produce the externally observed electrical (and thermal) characteristics of a PCM device. We address the first of these tasks (and introduce a vehicle for the second) by studying the formation of fcc polycrystalline grains from the as-deposited amorphous state in undoped Ge2Sb2Te5. We perform ex situ transmission electron microscopy membrane experiments and then match these observations against numerical simulation. Ramped-anneal experiments show that the temperature ramp-rate strongly influences the median grain size. By...

Sub-30nm scaling and high-speed operation of fully-confined Access-Devices for 3D crosspoint memory based on mixed-ionic-electronic-conduction (MIEC) materials

BEOL-friendly Access Devices (AD) based on Cu-containing MIEC materials [1-3] are shown to scale to the <;30nm CDs and <;12nm thicknesses found in advanced technology nodes. Switching speeds at the high (>100uA) currents of NVM writes can reach 15ns; NVM reads at typical (~5uA) current levels can be ≪1usec.