Neural networks are largely based on matrix computations. During forward inference, the most heavily used compute kernel is the matrix-vector multiplication (MVM): $W \vec{x} $. Inference is a first frontier for the deployment of next-generation hardware for neural network applications, as it is more readily deployed in edge devices, such as mobile devices or embedded processors with size, weight, and power constraints. Inference is also easier to implement in analog systems than training, which has more stringent device requirements. The main processing kernel used during inference is the MVM.
We demonstrate SONOS (silicon-oxide-nitride-oxide-silicon) analog memory arrays that are optimized for neural network inference. The devices are fabricated in a 40nm process and operated in the subthreshold regime for in-memory matrix multiplication. Subthreshold operation enables low conductances to be implemented with low error, which matches the typical weight distribution of neural networks, which is heavily skewed toward near-zero values. This leads to high accuracy in the presence of programming errors and process variations. We simulate the end-To-end neural network inference accuracy, accounting for the measured programming error, read noise, and retention loss in a fabricated SONOS array. Evaluated on the ImageNet dataset using ResNet50, the accuracy using a SONOS system is within 2.16% of floating-point accuracy without any retraining. The unique error properties and high On/Off ratio of the SONOS device allow scaling to large arrays without bit slicing, and enable an inference architecture that achieves 20 TOPS/W on ResNet50, a > 10× gain in energy efficiency over state-of-The-Art digital and analog inference accelerators.
To support the increasing demands for efficient deep neural network processing, accelerators based on analog in-memory computation of matrix multiplication have recently gained significant attention for reducing the energy of neural network inference. However, analog processing within memory arrays must contend with the issue of parasitic voltage drops across the metal interconnects, which distort the results of the computation and limit the array size. This work analyzes how parasitic resistance affects the end-to-end inference accuracy of state-of-the-art convolutional neural networks, and comprehensively studies how various design decisions at the device, circuit, architecture, and algorithm levels affect the system's sensitivity to parasitic resistance effects. A set of guidelines are provided for how to design analog accelerator hardware that is intrinsically robust to parasitic resistance, without any explicit compensation or re-training of the network parameters.
This project aimed to identify the performance-limiting mechanisms in mid- to far infrared (IR) sensors by probing photogenerated free carrier dynamics in model detector materials using scanning ultrafast electron microscopy (SUEM). SUEM is a recently developed method based on using ultrafast electron pulses in combination with optical excitations in a pump- probe configuration to examine charge dynamics with high spatial and temporal resolution and without the need for microfabrication. Five material systems were examined using SUEM in this project: polycrystalline lead zirconium titanate (a pyroelectric), polycrystalline vanadium dioxide (a bolometric material), GaAs (near IR), InAs (mid IR), and Si/SiO 2 system as a prototypical system for interface charge dynamics. The report provides detailed results for the Si/SiO 2 and the lead zirconium titanate systems.
Automated vehicles (AV) hold great promise for improving safety, as well as reducing congestion and emissions. In order to make automated vehicles commercially viable, a reliable and highperformance vehicle-based computing platform that meets ever-increasing computational demands will be key. Given the state of existing digital computing technology, designers will face significant challenges in meeting the needs of highly automated vehicles without exceeding thermal constraints or consuming a large portion of the energy available on vehicles, thus reducing range between charges or refills. The accompanying increases in energy for AV use will place increased demand on energy production and distribution infrastructure, which also motivates increasing computational energy efficiency.
Integration-technology feature shrink increases computing-system susceptibility to single-event effects (SEE). While modeling SEE faults will be critical, an integrated processor's scope makes physically correct modeling computationally intractable. Without useful models, presilicon evaluation of fault-tolerance approaches becomes impossible. To incorporate accurate transistor-level effects at a system scope, we present a multiscale simulation framework. Charge collection at the 1) device level determines 2) circuit-level transient duration and state-upset likelihood. Circuit effects, in turn, impact 3) register-transfer-level architecture-state corruption visible at 4) the system level. Thus, the physically accurate effects of SEEs in large-scale systems, executed on a high-performance computing (HPC) simulator, could be used to drive cross-layer radiation hardening by design. We demonstrate the capabilities of this model with two case studies. First, we determine a D flip-flop's sensitivity at the transistor level on 14-nm FinFet technology, validating the model against published cross sections. Second, we track and estimate faults in a microprocessor without interlocked pipelined stages (MIPS) processor for Adams 90% worst case environment in an isotropic space environment.
This presentation concludes in situ computation enables new approaches to linear algebra problems which can be both more effective and more efficient as compared to conventional digital systems. Preconditioning is well-suited to analog computation due to the tolerance for approximate solutions. When combined with prior work on in situ MVM for scientific computing, analog preconditioning can enable significant speedups for important linear algebra applications.
Over the past decade as Moore's Law has slowed, the need for new forms of computation that can provide sustainable performance improvements has risen. A new method, called in situ computing, has shown great potential to accelerate matrix vector multiplication (MVM), an important kernel for a diverse range of applications from neural networks to scientific computing. Existing in situ accelerators for scientific computing, however, have a significant limitation: These accelerators provide no acceleration for preconditioning-A key bottleneck in linear solvers and in scientific computing workflows. This paper enables in situ acceleration for state-of-The-Art linear solvers by demonstrating how to use a new in situ matrix inversion accelerator for analog preconditioning. As existing techniques that enable high precision and scalability for in situ MVM are inapplicable to in situ matrix inversion, new techniques to compensate for circuit non-idealities are proposed. Additionally, a new approach to bit slicing that enables splitting operands across multiple devices without external digital logic is proposed. For scalability, this paper demonstrates how in situ matrix inversion kernels can work in tandem with existing domain decomposition techniques to accelerate the solutions of arbitrarily large linear systems. The analog kernel can be directly integrated into existing preconditioning workflows, leveraging several well-optimized numerical linear algebra tools to improve the behavior of the circuit. The result is an analog preconditioner that is more effective (up to 50% fewer iterations) than the widely used incomplete LU factorization preconditioner, ILU(0), while also reducing the energy and execution time of each approximate solve operation by 1025x and 105x respectively.
Analog hardware accelerators, which perform computation within a dense memory array, have the potential to overcome the major bottlenecks faced by digital hardware for data-heavy workloads such as deep learning. Exploiting the intrinsic computational advantages of memory arrays, however, has proven to be challenging principally due to the overhead imposed by the peripheral circuitry and due to the non-ideal properties of memory devices that play the role of the synapse. We review the existing implementations of these accelerators for deep supervised learning, organizing our discussion around the different levels of the accelerator design hierarchy, with an emphasis on circuits and architecture. We explore and consolidate the various approaches that have been proposed to address the critical challenges faced by analog accelerators, for both neural network inference and training, and highlight the key design trade-offs underlying these techniques.
Emerging memory devices, such as resistive crossbars, have the capacity to store large amounts of data in a single array. Acquiring the data stored in large-capacity crossbars in a sequential fashion can become a bottleneck. We present practical methods, based on sparse sampling, to quickly acquire sparse data stored on emerging memory devices that support the basic summation kernel, reducing the acquisition time from linear to sub-linear. The experimental results show that at least an order of magnitude improvement in acquisition time can be achieved when the data are sparse. In addition, we show that the energy cost associated with our approach is competitive to that of the sequential method.
Analog resistive memories promise to reduce the energy of neural networks by orders of magnitude. However, the write variability and write nonlinearity of current devices prevent neural networks from training to high accuracy. We present a novel periodic carry method that uses a positional number system to overcome this while maintaining the benefit of parallel analog matrix operations. We demonstrate how noisy, nonlinear TaOx devices that could only train to 80% accuracy on MNIST, can now reach 97% accuracy, only 1% away from an ideal numeric accuracy of 98%. On a file type dataset, the TaOx devices achieve ideal numeric accuracy. In addition, low noise, linear Li1-xCoO2 devices train to ideal numeric accuracies using periodic carry on both datasets.
Resistive memory crossbars can dramatically reduce the energy required to perform computations in neural algorithms by three orders of magnitude when compared to an optimized digital ASIC [1]. For data intensive applications, the computational energy is dominated by moving data between the processor, SRAM, and DRAM. Analog crossbars overcome this by allowing data to be processed directly at each memory element. Analog crossbars accelerate three key operations that are the bulk of the computation in a neural network as illustrated in Fig 1: vector matrix multiplies (VMM), matrix vector multiplies (MVM), and outer product rank 1 updates (OPU)[2]. For an NxN crossbar the energy for each operation scales as the number of memory elements O(N2) [2]. This is because the crossbar performs its entire computation in one step, charging all the capacitances only once. Thus the CV2 energy of the array scales as array size. This fundamentally better than trying to read or write a digital memory. Each row of any NxN digital memory must be accessed one at a time, resulting in N columns of length O(N) being charged N times, requiring O(N3) energy to read a digital memory. Thus an analog crossbar has a fundamental O(N) energy scaling advantage over a digital system. Furthermore, if the read operation is done at low voltage and is therefore noise limited, the read energy can even be independent of the crossbar size, O(1) [2].
Biological neural networks continue to inspire new developments in algorithms and microelectronic hardware to solve challenging data processing and classification problems. Here, we survey the history of neural-inspired and neuromorphic computing in order to examine the complex and intertwined trajectories of the mathematical theory and hardware developed in this field. Early research focused on adapting existing hardware to emulate the pattern recognition capabilities of living organisms. Contributions from psychologists, mathematicians, engineers, neuroscientists, and other professions were crucial to maturing the field from narrowly-tailored demonstrations to more generalizable systems capable of addressing difficult problem classes such as object detection and speech recognition. Algorithms that leverage fundamental principles found in neuroscience such as hierarchical structure, temporal integration, and robustness to error have been developed, and some of these approaches are achieving world-leading performance on particular data classification tasks. In addition, novel microelectronic hardware is being developed to perform logic and to serve as memory in neuromorphic computing systems with optimized system integration and improved energy efficiency. Key to such advancements was the incorporation of new discoveries in neuroscience research, the transition away from strict structural replication and towards the functional replication of neural systems, and the use of mathematical theory frameworks to guide algorithm and hardware developments.
Resistive memories enable dramatic energy reductions for neural algorithms. We propose a general purpose neural architecture that can accelerate many different algorithms and determine the device properties that will be needed to run backpropagation on the neural architecture. To maintain high accuracy, the read noise standard deviation should be less than 5% of the weight range. The write noise standard deviation should be less than 0.4% of the weight range and up to 300% of a characteristic update (for the datasets tested). Asymmetric nonlinearities in the change in conductance vs pulse cause weight decay and significantly reduce the accuracy, while moderate symmetric nonlinearities do not have an effect. In order to allow for parallel reads and writes the write current should be less than 100 nA as well.
We present a study of the 'snap-back' regime of resistive switching hysteresis in bipolar TaOx memristors, identifying power signatures in the electronic transport. Using a simple model based on the thermal and electric field acceleration of ionic mobilities, we provide evidence that the 'snap-back' transition represents a crossover from a coupled thermal and electric-field regime to a primarily thermal regime, and is dictated by the reconnection of a ruptured conducting filament. We discuss how these power signatures can be used to limit filament radius growth, which is important for operational properties such as power, speed, and retention.
The exponential increase in data over the last decade presents a significant challenge to analytics efforts that seek to process and interpret such data for various applications. Neural-inspired computing approaches are being developed in order to leverage the computational properties of the analog, low-power data processing observed in biological systems. Analog resistive memory crossbars can perform a parallel read or a vector-matrix multiplication as well as a parallel write or a rank-1 update with high computational efficiency. For an N × N crossbar, these two kernels can be O(N) more energy efficient than a conventional digital memory-based architecture. If the read operation is noise limited, the energy to read a column can be independent of the crossbar size (O(1)). These two kernels form the basis of many neuromorphic algorithms such as image, text, and speech recognition. For instance, these kernels can be applied to a neural sparse coding algorithm to give an O(N) reduction in energy for the entire algorithm when run with finite precision. Sparse coding is a rich problem with a host of applications including computer vision, object tracking, and more generally unsupervised learning.
Final report for Cognitive Computing for Security LDRD 165613. It reports on the development of hybrid of general purpose/ne uromorphic computer architecture, with an emphasis on potential implementation with memristors.