The ASC program seeks to use machine learning to improve efficiencies in its stockpile stewardship mission. Moreover, there is a growing market for technologies dedicated to accelerating AI workloads. Many of these emerging architectures promise to provide savings in energy efficiency, area, and latency when compared to traditional CPUs for these types of applications — neuromorphic analog and digital technologies provide both low-power and configurable acceleration of challenging artificial intelligence (AI) algorithms. If designed into a heterogeneous system with other accelerators and conventional compute nodes, these technologies have the potential to augment the capabilities of traditional High Performance Computing (HPC) platforms [5]. This expanded computation space requires not only a new approach to physics simulation, but the ability to evaluate and analyze next-generation architectures specialized for AI/ML workloads in both traditional HPC and embedded ND applications. Developing this capability will enable ASC to understand how this hardware performs in both HPC and ND environments, improve our ability to port our applications, guide the development of computing hardware, and inform vendor interactions, leading them toward solutions that address ASC’s unique requirements.
This project evaluated the use of emerging spintronic memory devices for robust and efficient variational inference schemes. Variational inference (VI) schemes, which constrain the distribution for each weight to be a Gaussian distribution with a mean and standard deviation, are a tractable method for calculating posterior distributions of weights in a Bayesian neural network such that this neural network can also be trained using the powerful backpropagation algorithm. Our project focuses on domain-wall magnetic tunnel junctions (DW-MTJs), a powerful multi-functional spintronic synapse design that can achieve low power switching while also opening the pathway towards repeatable, analog operation using fabricated notches. Our initial efforts to employ DW-MTJs as an all-in-one stochastic synapse with both a mean and standard deviation didn’t end up meeting the quality metrics for hardware-friendly VI. In the future, new device stacks and methods for expressive anisotropy modification may make this idea still possible. However, as a fall back that immediately satisfies our requirements, we invented and detailed how the combination of a DW-MTJ synapse encoding the mean and a probabilistic Bayes-MTJ device, programmed via a ferroelectric or ionically modifiable layer, can robustly and expressively implement VI. This design includes a physics-informed small circuit model, that was scaled up to perform and demonstrate rigorous uncertainty quantification applications, up to and including small convolutional networks on a grayscale image classification task, and larger (Residual) networks implementing multi-channel image classification. Lastly, as these results and ideas all depend upon the idea of an inference application where weights (spintronic memory states) remain non-volatile, the retention of these synapses for the notched case was further interrogated. These investigations revealed and emphasized the importance of both notch geometry and anisotropy modification in order to further enhance the endurance of written spintronic states. In the near future, these results will be mapped to effective predictions for room temperature and elevated operation DW-MTJ memory retention, and experimentally verified when devices become available.
Analog computing has been widely proposed to improve the energy efficiency of multiple important workloads including neural network operations, and other linear algebra kernels. To properly evaluate analog computing and explore more complex workloads such as systems consisting of multiple analog data paths, system level simulations are required. Moreover, prior work on system architectures for analog computing often rely on custom simulators creating signficant additional design effort and complicating comparisons between different systems. To remedy these issues, this report describes the design and implementation of a flexible tile-based analog accelerator element for the Structural Simulation Toolkit (SST). The element focuses on heavily on the tile controller—an often neglected aspect of prior work—that is sufficiently versatile to simulate a wide range of different tile operations including neural network layers, signal processing kernels, and generic linear algebra operations without major constraints. The tile model also interoperates with existing SST memory and network models to reduce the overall development load and enable future simulation of heterogeneous systems with both conventional digital logic and analog compute tiles. Finally, both the tile and array models are designed to easily support future extensions as new analog operations and applications that can benefit from analog computing are developed.
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.
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.
We evaluate the sensitivity of neuromorphic inference accelerators based on silicon-oxide-nitride-oxide-silicon (SONOS) charge trap memory arrays to total ionizing dose (TID) effects. Data retention statistics were collected for 16 Mbit of 40-nm SONOS digital memory exposed to ionizing radiation from a Co-60 source, showing good retention of the bits up to the maximum dose of 500 krad(Si). Using this data, we formulate a rate-equation-based model for the TID response of trapped charge carriers in the ONO stack and predict the effect of TID on intermediate device states between 'program' and 'erase.' This model is then used to simulate arrays of low-power, analog SONOS devices that store 8-bit neural network weights and support in situ matrix-vector multiplication. We evaluate the accuracy of the irradiated SONOS-based inference accelerator on two image recognition tasks - CIFAR-10 and the challenging ImageNet data set - using state-of-the-art convolutional neural networks, such as ResNet-50. We find that across the data sets and neural networks evaluated, the accelerator tolerates a maximum TID between 10 and 100 krad(Si), with deeper networks being more susceptible to accuracy losses due to TID.
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.