Memristor-enabled neuromorphic computing systems provide a fast and energy-efficient approach to training neural networks1–4. However, convolutional neural networks (CNNs)—one of the most important models for image recognition5—have not yet been fully hardware-implemented using memristor crossbars, which are cross-point arrays with a memristor device at each intersection. Moreover, achieving software-comparable results is highly challenging owing to the poor yield, large variation and other non-ideal characteristics of devices6–9. Here we report the fabrication of high-yield, high-performance and uniform memristor crossbar arrays for the implementation of CNNs, which integrate eight 2,048-cell memristor arrays to improve parallel-computing efficiency. In addition, we propose an effective hybrid-training method to adapt to device imperfections and improve the overall system performance. We built a five-layer memristor-based CNN to perform MNIST10 image recognition, and achieved a high accuracy of more than 96 per cent. In addition to parallel convolutions using different kernels with shared inputs, replication of multiple identical kernels in memristor arrays was demonstrated for processing different inputs in parallel. The memristor-based CNN neuromorphic system has an energy efficiency more than two orders of magnitude greater than that of state-of-the-art graphics-processing units, and is shown to be scalable to larger networks, such as residual neural networks. Our results are expected to enable a viable memristor-based non-von Neumann hardware solution for deep neural networks and edge computing. A fully hardware-based memristor convolutional neural network using a hybrid training method achieves an energy efficiency more than two orders of magnitude greater than that of graphics-processing units.

Fully hardware-implemented memristor convolutional neural network

With their working mechanisms based on ion migration, the switching dynamics and electrical behaviour of memristive devices resemble those of synapses and neurons, making these devices promising candidates for brain-inspired computing. Built into large-scale crossbar arrays to form neural networks, they perform efficient in-memory computing with massive parallelism by directly using physical laws. The dynamical interactions between artificial synapses and neurons equip the networks with both supervised and unsupervised learning capabilities. Moreover, their ability to interface with analogue signals from sensors without analogue/digital conversions reduces the processing time and energy overhead. Although numerous simulations have indicated the potential of these networks for brain-inspired computing, experimental implementation of large-scale memristive arrays is still in its infancy. This Review looks at the progress, challenges and possible solutions for efficient brain-inspired computation with memristive implementations, both as accelerators for deep learning and as building blocks for spiking neural networks.

Memristive crossbar arrays for brain-inspired computing

This comprehensive review summarizes state of the art, challenges, and prospects of the neuro-inspired computing with emerging nonvolatile memory devices. First, we discuss the demand for developing neuro-inspired architecture beyond today’s von-Neumann architecture. Second, we summarize the various approaches to designing the neuromorphic hardware (digital versus analog, spiking versus nonspiking, online training versus offline training) and discuss why emerging nonvolatile memory is attractive for implementing the synapses in the neural network. Then, we discuss the desired device characteristics of the synaptic devices (e.g., multilevel states, weight update nonlinearity/asymmetry, variation/noise), and survey a few representative material systems and device prototypes reported in the literature that show the analog conductance tuning. These candidates include phase change memory, resistive memory, ferroelectric memory, floating-gate transistors, etc. Next, we introduce the crossbar array architecture to accelerate the weighted sum and weight update operations that are commonly used in the neuro-inspired machine learning algorithms, and review the recent progresses of array-level experimental demonstrations for pattern recognition tasks. In addition, we discuss the peripheral neuron circuit design issues and present a device-circuit-algorithm codesign methodology to evaluate the impact of nonideal device effects on the system-level performance (e.g., learning accuracy). Finally, we give an outlook on the customization of the learning algorithms for efficient hardware implementation.

Neuro-Inspired Computing With Emerging Nonvolatile Memorys

von Neumann architecture based computers isolate computation and storage (i.e. data is shuttled between computation blocks (processor) and memory blocks). The to-and-fro movement of data leads to a fundamental limitation of modern computers, known as the Memory wall. Logic in-Memory (LIM)/In-Memory Computing (IMC) approaches aim to address this bottleneck by directly computing inside memory units thereby eliminating energy-intensive and time-consuming data movement. Several recent works in literature, propose realization of logic function(s) directly using arrays of emerging resistive memory devices (example- memristors, RRAM/ReRAM, PCM, CBRAM, OxRAM, STT-MRAM etc.), rather than using conventional transistors for computing. The logic/embedded-side of digital systems (like processors, micro-controllers) can greatly benefit from such LIM realizations. However, the pure storage-side of digital systems (example SSDs, enterprise storage etc.) will not benefit much from such LIM approaches as when memory arrays are used for logic they lose their core functionality of storage. Thus, there is the need for an approach complementary to existing LIM techniques, that's more beneficial for the storage-side of digital systems; one that gives compute capability to memory arrays not at the cost of their existing stored states. Fundamentally, this would require memory nanodevice arrays that are capable of storing and computing simultaneously. In this paper, we propose a novel 'Simultaneous Logic in-Memory' (SLIM) methodology which is complementary to existing LIM approaches in literature. Through extensive experiments we demonstrate novel SLIM bitcells (1T-1R/2T-1R) comprising non-filamentary bilayer analog OxRAM devices with NMOS transistors. Proposed bitcells are capable of implementing both Memory and Logic operations simultaneously. Detailed programming scheme, array level implementation, and controller architecture are also proposed. Furthermore, to study the impact of proposed SLIM approach for real-world implementations, we performed analysis for two applications: (i) Sobel Edge Detection, and (ii) Binary Neural Network- Multi layer Perceptron (BNN-MLP). By performing all computations in SLIM bitcell array, huge Energy Delay Product (EDP) savings of ≈75× for 1T-1R (≈40× for 2T-1R) SLIM bitcell were observed for edge-detection application while EDP savings of ≈3.5× for 1T-1R (≈1.6× for 2T-1R) SLIM bitcell were observed for BNN-MLP application respectively, in comparison to conventional computing. EDP savings owing to reduction in data transfer between CPU ↔ memory is observed to be ≈780× (for both SLIM bitcells).

/pdf/slim-simultaneous-logic-in-memory-computing-exploiting-48akr75xhf.pdf

SLIM: Simultaneous Logic-in-Memory Computing Exploiting Bilayer Analog OxRAM Devices.

Domain-specific hardware is becoming a promising topic in the backdrop of improvement slow down for general-purpose processors due to the foreseeable end of Moore’s Law. Machine learning, especially deep neural networks (DNNs), has become the most dazzling domain witnessing successful applications in a wide spectrum of artificial intelligence (AI) tasks. The incomparable accuracy of DNNs is achieved by paying the cost of hungry memory consumption and high computational complexity, which greatly impedes their deployment in embedded systems. Therefore, the DNN compression concept was naturally proposed and widely used for memory saving and compute acceleration. In the past few years, a tremendous number of compression techniques have sprung up to pursue a satisfactory tradeoff between processing efficiency and application accuracy. Recently, this wave has spread to the design of neural network accelerators for gaining extremely high performance. However, the amount of related works is incredibly huge and the reported approaches are quite divergent. This research chaos motivates us to provide a comprehensive survey on the recent advances toward the goal of efficient compression and execution of DNNs without significantly compromising accuracy, involving both the high-level algorithms and their applications in hardware design. In this article, we review the mainstream compression approaches such as compact model, tensor decomposition, data quantization, and network sparsification. We explain their compression principles, evaluation metrics, sensitivity analysis, and joint-way use. Then, we answer the question of how to leverage these methods in the design of neural network accelerators and present the state-of-the-art hardware architectures. In the end, we discuss several existing issues such as fair comparison, testing workloads, automatic compression, influence on security, and framework/hardware-level support, and give promising topics in this field and the possible challenges as well. This article attempts to enable readers to quickly build up a big picture of neural network compression and acceleration, clearly evaluate various methods, and confidently get started in the right way.

Model Compression and Hardware Acceleration for Neural Networks: A Comprehensive Survey

Neuro-inspired architectures based on synaptic memory arrays have been proposed for on-chip acceleration of weighted sum and weight update in machine/deep learning algorithms. In this paper, we developed NeuroSim, a circuit-level macro model that estimates the area, latency, dynamic energy, and leakage power to facilitate the design space exploration of neuro-inspired architectures with mainstream and emerging device technologies. NeuroSim provides flexible interface and a wide variety of design options at the circuit and device level. Therefore, NeuroSim can be used by neural networks (NNs) as a supporting tool to provide circuit-level performance evaluation. With NeuroSim, an integrated framework can be built with hierarchical organization from the device level (synaptic device properties) to the circuit level (array architectures) and then to the algorithm level (NN topology), enabling instruction-accurate evaluation on the learning accuracy as well as the circuit-level performance metrics at the run-time of online learning. Using multilayer perceptron as a case-study algorithm, we investigated the impact of the “analog” emerging nonvolatile memory (eNVM)’s “nonideal” device properties and benchmarked the tradeoffs between SRAM, digital, and analog eNVM-based architectures for online learning and offline classification.

NeuroSim: A Circuit-Level Macro Model for Benchmarking Neuro-Inspired Architectures in Online Learning

NeuroSim+ is an integrated simulation framework for benchmarking synaptic devices and array architectures in terms of the system-level learning accuracy and hardware performance metrics. It has a hierarchical organization from the device level (transistor technology and memory cell models) to the circuit level (synaptic array architectures and neuron periphery) and then to the algorithm level (neural network topologies). In this work, we study the impact of the “analog” eNVM non-ideal device properties and benchmark the trade-offs of SRAM, digital and analog eNVM based array architectures for online learning and offline classification. The source code of NeuroSim+ version 1.0 is publicly available at https ://github. co m/neuro sim/MLP Neuro Sim.

NeuroSim+: An integrated device-to-algorithm framework for benchmarking synaptic devices and array architectures

Recent advances in deep learning have shown that Binary Neural Networks (BNNs) are capable of providing a satisfying accuracy on various image datasets with significant reduction in computation and memory cost. With both weights and activations binarized to +1 or −1 in BNNs, the high-precision multiply-and-accumulate (MAC) operations can be replaced by XNOR and bit-counting operations. In this work, we propose a RRAM synaptic architecture (XNOR-RRAM) with a bit-cell design of complementary word lines that implements equivalent XNOR and bit-counting operation in a parallel fashion. For large-scale matrices in fully connected layers or when the convolution kernels are unrolled in multiple channels, the array partition is necessary. Multi-level sense amplifiers (MLSAs) are employed as the intermediate interface for accumulating partial weighted sum. However, a low bit-level MLSA and intrinsic offset of MLSA may degrade the classification accuracy. We investigate the impact of sensing offsets on classification accuracy and analyze various design options with different sub-array sizes and sensing bit-levels. Experimental results with RRAM models and 65nm CMOS PDK show that the system with 128×128 sub-array size and 3-bit MLSA can achieve accuracies of 98.43% for MLP on MNIST and 86.08% for CNN on CIFAR-10, showing 0.34% and 2.39% degradation respectively compared to the accuracies of ideal BNN algorithms. The projected energy-efficiency of XNOR-RRAM is 141.18 TOPS/W, showing ∼33X improvement compared to the conventional RRAM synaptic architecture with sequential row-by-row read-out.

XNOR-RRAM: A scalable and parallel resistive synaptic architecture for binary neural networks

DNN+NeuroSim is an integrated framework to benchmark compute-in-memory (CIM) accelerators for deep neural networks, with hierarchical design options from device-level, to circuit-level and up to algorithm-level. A python wrapper is developed to interface NeuroSim with popular machine learning platforms such as Pytorch and Tensorflow. The framework supports automatic algorithm to hardware mapping, and evaluates both chip-level performance and inference accuracy with hardware constraints. In this work, we analyze the impact of reliability in "analog" synaptic devices, and analog-to-digital converter (ADC) quantization effects on the inference accuracy. Then we benchmark CIM accelerators based on SRAM and versatile emerging devices including RRAM, PCM, FeFET and ECRAM, from VGG to ResNet, and from CIFAR to ImageNet dataset, revealing the benefits of high on-state resistance, e.g. by using three-terminal synapses. The open-source code of DNN+NeuroSim is available at https://github.com/neurosim/DNN_NeuroSim_V1.0.

DNN+NeuroSim: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators with Versatile Device Technologies

The conductance tuning linearity is an important parameter of analog RRAM for neuromorphic computing. This work presents a novel methodology to improve the conductance tuning linearity of the filamentary RRAM. An electro-thermal modulation layer is designed and introduced to control the distribution of electric field and temperature in the filament region. For the first time, a HfO x based RRAM is demonstrated with linear analog SET, linear analog RESET, 50ns speed, 10× analog tuning window, 100kΩ on-state resistance, and high temperature retention for multilevel states. The excellent performances of the analog RRAM devices enable high accuracy online learning in a neural network.

Xiaochen Peng

Papers

NeuroSim: A Circuit-Level Macro Model for Benchmarking Neuro-Inspired Architectures in Online Learning

NeuroSim+: An integrated device-to-algorithm framework for benchmarking synaptic devices and array architectures

XNOR-RRAM: A scalable and parallel resistive synaptic architecture for binary neural networks

DNN+NeuroSim: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators with Versatile Device Technologies

A Methodology to Improve Linearity of Analog RRAM for Neuromorphic Computing