Search

Current state-of-the-art employs approximate multipliers to address the highly increased power demands of DNN accelerators. However, evaluating the accuracy of approximate DNNs is cumbersome due to the lack of adequate support for approximate arithmetic in DNN frameworks. We address this inefficiency by presenting AdaPT, a fast emulation framework that extends PyTorch to support approximate inference as well as approximation-aware retraining. AdaPT can be seamlessly deployed and is compatible with the most DNNs. We evaluate the framework on several DNN models and application fields including CNNs, LSTMs, and GANs for a number of approximate multipliers with distinct bitwidth values. The results show substantial error recovery from approximate re-training and reduced inference time up to 53.9x with respect to the baseline approximate implementation., Accepted for publication in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems

Recently, automated co-design of machine learning (ML) models and accelerator architectures has attracted significant attention from both the industry and academia. However, most co-design frameworks either explore a limited search space or employ suboptimal exploration techniques for simultaneous design decision investigations of the ML model and the accelerator. Furthermore, training the ML model and simulating the accelerator performance is computationally expensive. To address these limitations, this work proposes a novel neural architecture and hardware accelerator co-design framework, called CODEBench. It is composed of two new benchmarking sub-frameworks, CNNBench and AccelBench, which explore expanded design spaces of convolutional neural networks (CNNs) and CNN accelerators. CNNBench leverages an advanced search technique, BOSHNAS, to efficiently train a neural heteroscedastic surrogate model to converge to an optimal CNN architecture by employing second-order gradients. AccelBench performs cycle-accurate simulations for a diverse set of accelerator architectures in a vast design space. With the proposed co-design method, called BOSHCODE, our best CNN-accelerator pair achieves 1.4% higher accuracy on the CIFAR-10 dataset compared to the state-of-the-art pair, while enabling 59.1% lower latency and 60.8% lower energy consumption. On the ImageNet dataset, it achieves 3.7% higher Top1 accuracy at 43.8% lower latency and 11.2% lower energy consumption. CODEBench outperforms the state-of-the-art framework, i.e., Auto-NBA, by achieving 1.5% higher accuracy and 34.7x higher throughput, while enabling 11.0x lower energy-delay product (EDP) and 4.0x lower chip area on CIFAR-10., Comment: Published at ACM Transactions on Embedded Computing Systems. Code available at https://github.com/jha-lab/codebench

Recent efforts for improving the performance of neural network (NN) accelerators that meet today's application requirements have given rise to a new trend of logic-based NN inference relying on fixed function combinational logic. Mapping such large Boolean functions with many input variables and product terms to digital signal processors (DSPs) on Field-programmable gate arrays (FPGAs) needs a novel framework considering the structure and the reconfigurability of DSP blocks during this process. The proposed methodology in this paper maps the fixed function combinational logic blocks to a set of Boolean functions where Boolean operations corresponding to each function are mapped to DSP devices rather than look-up tables (LUTs) on the FPGAs to take advantage of the high performance, low latency, and parallelism of DSP blocks. % This paper also presents an innovative design and optimization methodology for compilation and mapping of NNs, utilizing fixed function combinational logic to DSPs on FPGAs employing high-level synthesis flow. % Our experimental evaluations across several \REVone{datasets} and selected NNs demonstrate the comparable performance of our framework in terms of the inference latency and output accuracy compared to prior art FPGA-based NN accelerators employing DSPs., Comment: 25 page, 10 figures. Under review

K-nearest neighbor search is one of the fundamental tasks in various applications and the hierarchical navigable small world (HNSW) has recently drawn attention in large-scale cloud services, as it easily scales up the database while offering fast search. On the other hand, a computational storage device (CSD) that combines programmable logic and storage modules on a single board becomes popular to address the data bandwidth bottleneck of modern computing systems. In this paper, we propose a computational storage platform that can accelerate a large-scale graph-based nearest neighbor search algorithm based on SmartSSD CSD. To this end, we modify the algorithm more amenable on the hardware and implement two types of accelerators using HLS- and RTL-based methodology with various optimization methods. In addition, we scale up the proposed platform to have 4 SmartSSDs and apply graph parallelism to boost the system performance further. As a result, the proposed computational storage platform achieves 75.59 query per second throughput for the SIFT1B dataset at 258.66W power dissipation, which is 12.83x and 17.91x faster and 10.43x and 24.33x more energy efficient than the conventional CPU-based and GPU-based server platform, respectively. With multi-terabyte storage and custom acceleration capability, we believe that the proposed computational storage platform is a promising solution for cost-sensitive cloud datacenters., Extension of FCCM 20201 and Accepted in Transaction on Computers

In this work, we demonstrate the offline FPGA realization of both recurrent and feedforward neural network (NN)-based equalizers for nonlinearity compensation in coherent optical transmission systems. First, we present a realization pipeline showing the conversion of the models from Python libraries to the FPGA chip synthesis and implementation. Then, we review the main alternatives for the hardware implementation of nonlinear activation functions. The main results are divided into three parts: a performance comparison, an analysis of how activation functions are implemented, and a report on the complexity of the hardware. The performance in Q-factor is presented for the cases of bidirectional long-short-term memory coupled with convolutional NN (biLSTM + CNN) equalizer, CNN equalizer, and standard 1-StpS digital back-propagation (DBP) for the simulation and experiment propagation of a single channel dual-polarization (SC-DP) 16QAM at 34 GBd along 17x70km of LEAF. The biLSTM+CNN equalizer provides a similar result to DBP and a 1.7 dB Q-factor gain compared with the chromatic dispersion compensation baseline in the experimental dataset. After that, we assess the Q-factor and the impact of hardware utilization when approximating the activation functions of NN using Taylor series, piecewise linear, and look-up table (LUT) approximations. We also show how to mitigate the approximation errors with extra training and provide some insights into possible gradient problems in the LUT approximation. Finally, to evaluate the complexity of hardware implementation to achieve 200G and 400G throughput, fixed-point NN-based equalizers with approximated activation functions are developed and implemented in an FPGA., Comment: Invited paper at Journal of Lightwave Technology - IEEE

Training deep neural networks (DNNs) is a computationally expensive job, which can take weeks or months even with high performance GPUs. As a remedy for this challenge, community has started exploring the use of more efficient data representations in the training process, e.g., block floating point (BFP). However, prior work on BFP-based DNN accelerators rely on a specific BFP representation making them less versatile. This paper builds upon an algorithmic observation that we can accelerate the training by leveraging multiple BFP precisions without compromising the finally achieved accuracy. Backed up by this algorithmic opportunity, we develop a flexible DNN training accelerator, dubbed FlexBlock, which supports three different BFP precision modes, possibly different among activation, weight, and gradient tensors. While several prior works proposed such multi-precision support for DNN accelerators, not only do they focus only on the inference, but also their core utilization is suboptimal at a fixed precision and specific layer types when the training is considered. Instead, FlexBlock is designed in such a way that high core utilization is achievable for i) various layer types, and ii) three BFP precisions by mapping data in a hierarchical manner to its compute units. We evaluate the effectiveness of FlexBlock architecture using well-known DNNs on CIFAR, ImageNet and WMT14 datasets. As a result, training in FlexBlock significantly improves the training speed by 1.5~5.3x and the energy efficiency by 2.4~7.0x on average compared to other training accelerators and incurs marginal accuracy loss compared to full-precision training.

Analog computing is attractive compared to digital computing due to its potential for achieving higher computational density and higher energy efficiency. However, unlike digital circuits, conventional analog computing circuits cannot be easily mapped across different process nodes due to differences in transistor biasing regimes, temperature variations and limited dynamic range. In this work, we generalize the previously reported margin-propagation-based analog computing framework for designing novel \textit{shape-based analog computing} (S-AC) circuits that can be easily cross-mapped across different process nodes. Similar to digital designs S-AC designs can also be scaled for precision, speed, and power. As a proof-of-concept, we show several examples of S-AC circuits implementing mathematical functions that are commonly used in machine learning (ML) architectures. Using circuit simulations we demonstrate that the circuit input/output characteristics remain robust when mapped from a planar CMOS 180nm process to a FinFET 7nm process. Also, using benchmark datasets we demonstrate that the classification accuracy of a S-AC based neural network remains robust when mapped across the two processes and to changes in temperature., Comment: 14 Pages, 15 Figures, 5 Tables. This work has been accepted in IEEE for publication. Copyright may be transferred without notice, after which this version may no longer be accessible

Deep Learning (DL) has transformed the automation of a wide range of industries and finds increasing ubiquity in society. The high complexity of DL models and its widespread adoption has led to global energy consumption doubling every 3-4 months. Currently, the relationship between the DL model configuration and energy consumption is not well established. At a general computational energy model level, there is both strong dependency to both the hardware architecture (e.g. generic processors with different configuration of inner components- CPU and GPU, programmable integrated circuits - FPGA), as well as different interacting energy consumption aspects (e.g., data movement, calculation, control). At the DL model level, we need to translate non-linear activation functions and its interaction with data into calculation tasks. Current methods mainly linearize nonlinear DL models to approximate its theoretical FLOPs and MACs as a proxy for energy consumption. Yet, this is inaccurate (est. 93\% accuracy) due to the highly nonlinear nature of many convolutional neural networks (CNNs) for example. In this paper, we develop a bottom-level Transistor Operations (TOs) method to expose the role of non-linear activation functions and neural network structure in energy consumption. We translate a range of feedforward and CNN models into ALU calculation tasks and then TO steps. This is then statistically linked to real energy consumption values via a regression model for different hardware configurations and data sets. We show that our proposed TOs method can achieve a 93.61% - 99.51% precision in predicting its energy consumption., 15 pages, 11 figures

Power delivery network (PDN) analysis and thermal analysis are computationally expensive tasks that are essential for successful IC design. Algorithmically, both these analyses have similar computational structure and complexity as they involve the solution to a partial differential equation of the same form. This paper converts these analyses into image-to-image and sequence-to-sequence translation tasks, which allows leveraging a class of machine learning models with an encoder-decoder-based generative (EDGe) architecture to address the time-intensive nature of these tasks. For PDN analysis, we propose two networks: (i) IREDGe: a full-chip static and dynamic IR drop predictor and (ii) EMEDGe: electromigration (EM) hotspot classifier based on input power, power grid distribution, and power pad distribution patterns. For thermal analysis, we propose ThermEDGe, a full-chip static and dynamic temperature estimator based on input power distribution patterns for thermal analysis. These networks are transferable across designs synthesized within the same technology and packing solution. The networks predict on-chip IR drop, EM hotspot locations, and temperature in milliseconds with negligibly small errors against commercial tools requiring several hours., Comment: 26 pages, 17 figures, Submitted to TODAES for review. arXiv admin note: text overlap with arXiv:2009.09009

Hyperdimensional computing (HDC) is an emerging computational framework inspired by the brain that operates on vectors with thousands of dimensions to emulate cognition. Unlike conventional computational frameworks that operate on numbers, HDC, like the brain, uses high dimensional random vectors and is capable of one-shot learning. HDC is based on a well-defined set of arithmetic operations and is highly error-resilient. The core operations of HDC manipulate HD vectors in bulk bit-wise fashion, offering many opportunities to leverage parallelism. Unfortunately, on conventional Von-Neuman architectures, the continuous movement of HD vectors among the processor and the memory can make the cognition task prohibitively slow and energy-intensive. Hardware accelerators only marginally improve related metrics. On the contrary, only partial implementation of an HDC framework inside memory, using emerging memristive devices, has reported considerable performance/energy gains. This paper presents an architecture based on racetrack memory (RTM) to conduct and accelerate the entire HDC framework within the memory. The proposed solution requires minimal additional CMOS circuitry and uses a read operation across multiple domains in RTMs called transverse read (TR) to realize exclusive-or (XOR) and addition operations. To minimize the overhead the CMOS circuitry, we propose an RTM nanowires-based counting mechanism that leverages the TR operation and the standard RTM operations. Using language recognition as the use case demonstrates 7.8x and 5.3x reduction in the overall runtime and energy consumption compared to the FPGA design, respectively. Compared to the state-of-the-art in-memory implementation, the proposed HDC system reduces the energy consumption by 8.6x., Preprint, accepted for publication, ACM Transactions on Embedded Computing Systems. ACM Trans. Embed. Comput. Syst. (March 2022)

The Transformer has been an indispensable staple in deep learning. However, for real-life applications, it is very challenging to deploy efficient Transformers due to immense parameters and operations of models. To relieve this burden, exploiting sparsity is an effective approach to accelerate Transformers. Newly emerging Ampere GPUs leverage a 2:4 sparsity pattern to achieve model acceleration, while it can hardly meet the diverse algorithm and hardware constraints when deploying models. By contrast, we propose an algorithm-hardware co-optimized framework to flexibly and efficiently accelerate Transformers by utilizing general N:M sparsity patterns. (1) From algorithm perspective, we propose a sparsity inheritance mechanism along with an inherited dynamic pruning (IDP) method to obtain a series of N:M sparse candidate Transformers rapidly. A model compression scheme is further proposed to significantly reduce the storage requirement for deployment. (2) From hardware perspective, we present a flexible and efficient hardware architecture, namely STA, to achieve significant speedup when deploying N:M sparse Transformers. STA features not only a computing engine unifying both sparse-dense and dense-dense matrix multiplications with high computational efficiency but also a scalable softmax module eliminating the latency from intermediate off-chip data communication. Experimental results show that compared to other methods, N:M sparse Transformers, generated using IDP, achieves an average of 6.7% improvement on accuracy with high training efficiency. Moreover, STA can achieve 14.47x and 11.33x speedup compared to Intel i9-9900X and NVIDIA RTX 2080 Ti, respectively, and perform 2.00-19.47x faster inference than the state-of-the-art FPGA-based accelerators for Transformers., To appear in IEEE Transactions on Very Large Scale Integration (VLSI) Systems

Neural networks (NNs) are growing in importance and complexity. A neural network's performance (and energy efficiency) can be bound either by computation or memory resources. The processing-in-memory (PIM) paradigm, where computation is placed near or within memory arrays, is a viable solution to accelerate memory-bound NNs. However, PIM architectures vary in form, where different PIM approaches lead to different trade-offs. Our goal is to analyze, discuss, and contrast DRAM-based PIM architectures for NN performance and energy efficiency. To do so, we analyze three state-of-the-art PIM architectures: (1) UPMEM, which integrates processors and DRAM arrays into a single 2D chip; (2) Mensa, a 3D-stack-based PIM architecture tailored for edge devices; and (3) SIMDRAM, which uses the analog principles of DRAM to execute bit-serial operations. Our analysis reveals that PIM greatly benefits memory-bound NNs: (1) UPMEM provides 23x the performance of a high-end GPU when the GPU requires memory oversubscription for a general matrix-vector multiplication kernel; (2) Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Google Edge TPU for 24 Google edge NN models; and (3) SIMDRAM outperforms a CPU/GPU by 16.7x/1.4x for three binary NNs. We conclude that the ideal PIM architecture for NN models depends on a model's distinct attributes, due to the inherent architectural design choices., This is an extended and updated version of a paper published in IEEE Micro, pp. 1-14, 29 Aug. 2022. arXiv admin note: text overlap with arXiv:2109.14320

Recent years have seen a phenomenal rise in performance and applications of transformer neural networks. The family of transformer networks, including Bidirectional Encoder Representations from Transformer (BERT), Generative Pretrained Transformer (GPT) and Vision Transformer (ViT), have shown their effectiveness across Natural Language Processing (NLP) and Computer Vision (CV) domains. Transformer-based networks such as ChatGPT have impacted the lives of common men. However, the quest for high predictive performance has led to an exponential increase in transformers' memory and compute footprint. Researchers have proposed techniques to optimize transformer inference at all levels of abstraction. This paper presents a comprehensive survey of techniques for optimizing the inference phase of transformer networks. We survey techniques such as knowledge distillation, pruning, quantization, neural architecture search and lightweight network design at the algorithmic level. We further review hardware-level optimization techniques and the design of novel hardware accelerators for transformers. We summarize the quantitative results on the number of parameters/FLOPs and accuracy of several models/techniques to showcase the tradeoff exercised by them. We also outline future directions in this rapidly evolving field of research. We believe that this survey will educate both novice and seasoned researchers and also spark a plethora of research efforts in this field.

Transformer networks have emerged as the state-of-the-art approach for natural language processing tasks and are gaining popularity in other domains such as computer vision and audio processing. However, the efficient hardware acceleration of transformer models poses new challenges due to their high arithmetic intensities, large memory requirements, and complex dataflow dependencies. In this work, we propose ITA, a novel accelerator architecture for transformers and related models that targets efficient inference on embedded systems by exploiting 8-bit quantization and an innovative softmax implementation that operates exclusively on integer values. By computing on-the-fly in streaming mode, our softmax implementation minimizes data movement and energy consumption. ITA achieves competitive energy efficiency with respect to state-of-the-art transformer accelerators with 16.9 TOPS/W, while outperforming them in area efficiency with 5.93 TOPS/mm$^2$ in 22 nm fully-depleted silicon-on-insulator technology at 0.8 V., Accepted for publication at the 2023 ACM/IEEE International Symposium on Low Power Electronics and Design (ISLPED)

Graph neural networks (GNNs) have emerged as a powerful approach for modelling and learning from graph-structured data. Multiple fields have since benefitted enormously from the capabilities of GNNs, such as recommendation systems, social network analysis, drug discovery, and robotics. However, accelerating and efficiently processing GNNs require a unique approach that goes beyond conventional artificial neural network accelerators, due to the substantial computational and memory requirements of GNNs. The slowdown of scaling in CMOS platforms also motivates a search for alternative implementation substrates. In this paper, we present GHOST, the first silicon-photonic hardware accelerator for GNNs. GHOST efficiently alleviates the costs associated with both vertex-centric and edge-centric operations. It implements separately the three main stages involved in running GNNs in the optical domain, allowing it to be used for the inference of various widely used GNN models and architectures, such as graph convolution networks and graph attention networks. Our simulation studies indicate that GHOST exhibits at least 10.2x better throughput and 3.8x better energy efficiency when compared to GPU, TPU, CPU and multiple state-of-the-art GNN hardware accelerators.

Graph Neural Networks (GNNs) are emerging as a powerful tool for learning from graph-structured data and performing sophisticated inference tasks in various application domains. Although GNNs have been shown to be effective on modest-sized graphs, training them on large-scale graphs remains a significant challenge due to lack of efficient data access and data movement methods. Existing frameworks for training GNNs use CPUs for graph sampling and feature aggregation, while the training and updating of model weights are executed on GPUs. However, our in-depth profiling shows the CPUs cannot achieve the throughput required to saturate GNN model training throughput, causing gross under-utilization of expensive GPU resources. Furthermore, when the graph and its embeddings do not fit in the CPU memory, the overhead introduced by the operating system, say for handling page-faults, comes in the critical path of execution. To address these issues, we propose the GPU Initiated Direct Storage Access (GIDS) dataloader, to enable GPU-oriented GNN training for large-scale graphs while efficiently utilizing all hardware resources, such as CPU memory, storage, and GPU memory with a hybrid data placement strategy. By enabling GPU threads to fetch feature vectors directly from storage, GIDS dataloader solves the memory capacity problem for GPU-oriented GNN training. Moreover, GIDS dataloader leverages GPU parallelism to tolerate storage latency and eliminates expensive page-fault overhead. Doing so enables us to design novel optimizations for exploiting locality and increasing effective bandwidth for GNN training. Our evaluation using a single GPU on terabyte-scale GNN datasets shows that GIDS dataloader accelerates the overall DGL GNN training pipeline by up to 392X when compared to the current, state-of-the-art DGL dataloader., Under Submission. Source code: https://github.com/jeongminpark417/GIDS

Quantization is commonly used in Deep Neural Networks (DNNs) to reduce the storage and computational complexity by decreasing the arithmetical precision of activations and weights, a.k.a. tensors. Efficient hardware architectures employ linear quantization to enable the deployment of recent DNNs onto embedded systems and mobile devices. However, linear uniform quantization cannot usually reduce the numerical precision to less than 8 bits without sacrificing high performance in terms of model accuracy. The performance loss is due to the fact that tensors do not follow uniform distributions. In this paper, we show that a significant amount of tensors fit into an exponential distribution. Then, we propose DNA-TEQ to exponentially quantize DNN tensors with an adaptive scheme that achieves the best trade-off between numerical precision and accuracy loss. The experimental results show that DNA-TEQ provides a much lower quantization bit-width compared to previous proposals, resulting in an average compression ratio of 40% over the linear INT8 baseline, with negligible accuracy loss and without retraining the DNNs. Besides, DNA-TEQ leads the way in performing dot-product operations in the exponential domain, which saves 66% of energy consumption on average for a set of widely used DNNs., 8 pages, 8 figures, 5 tables

Recent trends in deep learning (DL) imposed hardware accelerators as the most viable solution for several classes of high-performance computing (HPC) applications such as image classification, computer vision, and speech recognition. This survey summarizes and classifies the most recent advances in designing DL accelerators suitable to reach the performance requirements of HPC applications. In particular, it highlights the most advanced approaches to support deep learning accelerations including not only GPU and TPU-based accelerators but also design-specific hardware accelerators such as FPGA-based and ASIC-based accelerators, Neural Processing Units, open hardware RISC-V-based accelerators and co-processors. The survey also describes accelerators based on emerging memory technologies and computing paradigms, such as 3D-stacked Processor-In-Memory, non-volatile memories (mainly, Resistive RAM and Phase Change Memories) to implement in-memory computing, Neuromorphic Processing Units, and accelerators based on Multi-Chip Modules. The survey classifies the most influential architectures and technologies proposed in the last years, with the purpose of offering the reader a comprehensive perspective in the rapidly evolving field of deep learning. Finally, it provides some insights into future challenges in DL accelerators such as quantum accelerators and photonics., Preprint version of our manuscript submitted to the journal @ ACM CSUR (35 pages plus Appendix) on June 22nd, 2023

Graph neural networks (GNNs) are emerging for machine learning research on graph-structured data. GNNs achieve state-of-the-art performance on many tasks, but they face scalability challenges when it comes to real-world applications that have numerous data and strict latency requirements. Many studies have been conducted on how to accelerate GNNs in an effort to address these challenges. These acceleration techniques touch on various aspects of the GNN pipeline, from smart training and inference algorithms to efficient systems and customized hardware. As the amount of research on GNN acceleration has grown rapidly, there lacks a systematic treatment to provide a unified view and address the complexity of relevant works. In this survey, we provide a taxonomy of GNN acceleration, review the existing approaches, and suggest future research directions. Our taxonomic treatment of GNN acceleration connects the existing works and sets the stage for further development in this area.

To ensure resilient neural network processing on even unreliable hardware, comprehensive reliability analysis against various hardware faults is generally required before the deep neural network models are deployed, and efficient error injection tools are highly demanded. However, most existing fault injection tools remain rather limited to basic fault injection to neurons and fail to provide fine-grained vulnerability analysis capability. In addition, many of the fault injection tools still need to change the neural network models and make the fault injection closely coupled with normal neural network processing, which further complicates the use of the fault injection tools and slows down the fault simulation. In this work, we propose MRFI, a highly configurable multi-resolution fault injection tool for deep neural networks. It enables users to modify an independent fault configuration file rather than neural network models for the fault injection and vulnerability analysis. Particularly, it integrates extensive fault analysis functionalities from different perspectives and enables multi-resolution investigation of the vulnerability of neural networks. In addition, it does not modify the major neural network computing framework of PyTorch. Hence, it allows parallel processing on GPUs naturally and exhibits fast fault simulation according to our experiments., 8 pages, 11 figures, source code is on https://github.com/fffasttime/MRFI

Machine learning (ML) has become a prevalent approach to tame the complexity of design space exploration for domain-specific architectures. While appealing, using ML for design space exploration is challenging. First, it is not straightforward to identify the most suitable algorithms for a target problem space from an ever-increasing pool of ML methods. Second, the trade-offs such as performance vs. sample efficiency across the different ML-driven approaches is rather inconclusive. Finally, the lack of a holistic infrastructure for a fair and reproducible comparisons between these approaches hinders progress and impedes creating meaningful artifacts. To mitigate these challenges, we introduce ARCHitecture GYM (ARCHGYM), a framework that connects a variety of search algorithms to architecture simulators with the ability to extend with additional algorithms and environments as needed owning to its abstract interface. To demonstrate its utility, we evaluate multiple vanilla and domain-specific search algorithms in designing a custom memory controller, deep neural network accelerators, and a custom SoC for AR/VR workloads, collectively through more than 21, 000 experiments. The results from these studies indicate that, diverse ML algorithms are equally favorable to meet the user-defined target speficication if its hyperparameters are tuned correctly; no one search algorithm is necessarily better than another (e.g., reinforcement learning vs. Bayesian methods). We coin the term “hyperparameter lottery” to describe the relatively probable chance for a search algorithm to find an optimal design provided the right hyperparameter is chosen. We finally demonstrate how ARCHGYM enables data collection from ML algorithms to facilitate future research. We showcase its benefits by developing an ML-based proxy cost model with an RMSE 0.61%. This relatively high-accuracy cost model offers 2,000× faster simulation time., For artifact evaluation

Deep Neural Networks (DNNs) have demonstrated impressive performance across a wide range of tasks. However, deploying DNNs on edge devices poses significant challenges due to stringent power and computational budgets. An effective solution to this issue is software-hardware (SW-HW) co-design, which allows for the tailored creation of DNN models and hardware architectures that optimally utilize available resources. However, SW-HW co-design traditionally suffers from slow optimization speeds because their optimizers do not make use of heuristic knowledge, also known as the ``cold start'' problem. In this study, we present a novel approach that leverages Large Language Models (LLMs) to address this issue. By utilizing the abundant knowledge of pre-trained LLMs in the co-design optimization process, we effectively bypass the cold start problem, substantially accelerating the design process. The proposed method achieves a significant speedup of 25x. This advancement paves the way for the rapid and efficient deployment of DNNs on edge devices.

Dataflow scheduling decisions are of vital importance to neural network (NN) accelerators. Recent scalable NN accelerators support a rich set of advanced dataflow techniques. The problems of comprehensively representing and quickly finding optimized dataflow schemes thus become significantly more complicated and challenging. In this work, we first propose comprehensive and pragmatic dataflow representations for temporal and spatial scheduling on scalable multi-node NN architectures. An informal hierarchical taxonomy highlights the tight coupling across different levels of the dataflow space as the major difficulty for fast design exploration. A set of formal tensor-centric directives accurately express various inter-layer and intra-layer schemes, and allow for quickly determining their validity and efficiency. We then build a generic, optimized, and fast dataflow solver, KAPLA, which makes use of the pragmatic directives to explore the design space with effective validity check and efficiency estimation. KAPLA decouples the upper inter-layer level for fast pruning, and solves the lower intra-layer schemes with a novel bottom-up cost descending method. KAPLA achieves within only 2.2% and 7.7% energy overheads on the result dataflow for training and inference, respectively, compared to the exhaustively searched optimal schemes. It also outperforms random and machine-learning-based approaches, with more optimized results and orders of magnitude faster search speedup.

Emerging AI applications such as ChatGPT, graph convolutional networks, and other deep neural networks require massive computational resources for training and inference. Contemporary computing platforms such as CPUs, GPUs, and TPUs are struggling to keep up with the demands of these AI applications. Non-coherent optical computing represents a promising approach for light-speed acceleration of AI workloads. In this paper, we show how cross-layer design can overcome challenges in non-coherent optical computing platforms. We describe approaches for optical device engineering, tuning circuit enhancements, and architectural innovations to adapt optical computing to a variety of AI workloads. We also discuss techniques for hardware/software co-design that can intelligently map and adapt AI software to improve its performance on non-coherent optical computing platforms.

Tensor processing units (TPUs), specialized hardware accelerators for machine learning tasks, have shown significant performance improvements when executing convolutional layers in convolutional neural networks (CNNs). However, they struggle to maintain the same efficiency in fully connected (FC) layers, leading to suboptimal hardware utilization. In-memory analog computing (IMAC) architectures, on the other hand, have demonstrated notable speedup in executing FC layers. This paper introduces a novel, heterogeneous, mixed-signal, and mixed-precision architecture that integrates an IMAC unit with an edge TPU to enhance mobile CNN performance. To leverage the strengths of TPUs for convolutional layers and IMAC circuits for dense layers, we propose a unified learning algorithm that incorporates mixed-precision training techniques to mitigate potential accuracy drops when deploying models on the TPU-IMAC architecture. The simulations demonstrate that the TPU-IMAC configuration achieves up to $2.59\times$ performance improvements, and $88\%$ memory reductions compared to conventional TPU architectures for various CNN models while maintaining comparable accuracy. The TPU-IMAC architecture shows potential for various applications where energy efficiency and high performance are essential, such as edge computing and real-time processing in mobile devices. The unified training algorithm and the integration of IMAC and TPU architectures contribute to the potential impact of this research on the broader machine learning landscape.

Many hardware-aware neural architecture search (NAS) methods have been developed to optimize the topology of neural networks (NN) with the joint objectives of higher accuracy and lower latency. Recently, both accuracy and latency predictors have been used in NAS with great success, achieving high sample efficiency and accurate modeling of hardware (HW) device latency respectively. However, a new accuracy predictor needs to be trained for every new NAS search space or NN task, and a new latency predictor needs to be additionally trained for every new HW device. In this paper, we explore methods to enable multi-task, multi-search-space, and multi-HW adaptation of accuracy and latency predictors to reduce the cost of NAS. We introduce a novel search-space independent NN encoding based on zero-cost proxies that achieves sample-efficient prediction on multiple tasks and NAS search spaces, improving the end-to-end sample efficiency of latency and accuracy predictors by over an order of magnitude in multiple scenarios. For example, our NN encoding enables multi-search-space transfer of latency predictors from NASBench-201 to FBNet (and vice-versa) in under 85 HW measurements, a 400$\times$ improvement in sample efficiency compared to a recent meta-learning approach. Our method also improves the total sample efficiency of accuracy predictors by over an order of magnitude. Finally, we demonstrate the effectiveness of our method for multi-search-space and multi-task accuracy prediction on 28 NAS search spaces and tasks.

There is a need for machine learning models to evolve in unsupervised circumstances. New classifications may be introduced, unexpected faults may occur, or the initial dataset may be small compared to the data-points presented to the system during normal operation. Implementing such a system using neural networks involves significant mathematical complexity, which is a major issue in power-critical edge applications. This paper proposes a novel field-programmable gate-array infrastructure for online learning, implementing a low-complexity machine learning algorithm called the Tsetlin Machine. This infrastructure features a custom-designed architecture for run-time learning management, providing on-chip offline and online learning. Using this architecture, training can be carried out on-demand on the \ac{FPGA} with pre-classified data before inference takes place. Additionally, our architecture provisions online learning, where training can be interleaved with inference during operation. Tsetlin Machine (TM) training naturally descends to an optimum, with training also linked to a threshold hyper-parameter which is used to reduce the probability of issuing feedback as the TM becomes trained further. The proposed architecture is modular, allowing the data input source to be easily changed, whilst inbuilt cross-validation infrastructure allows for reliable and representative results during system testing. We present use cases for online learning using the proposed infrastructure and demonstrate the energy/performance/accuracy trade-offs.

Deployment of modern TinyML tasks on small battery-constrained IoT devices requires high computational energy efficiency. Analog In-Memory Computing (IMC) using non-volatile memory (NVM) promises major efficiency improvements in deep neural network (DNN) inference and serves as on-chip memory storage for DNN weights. However, IMC's functional flexibility limitations and their impact on performance, energy, and area efficiency are not yet fully understood at the system level. To target practical end-to-end IoT applications, IMC arrays must be enclosed in heterogeneous programmable systems, introducing new system-level challenges which we aim at addressing in this work. We present a heterogeneous tightly-coupled clustered architecture integrating 8 RISC-V cores, an in-memory computing accelerator (IMA), and digital accelerators. We benchmark the system on a highly heterogeneous workload such as the Bottleneck layer from a MobileNetV2, showing 11.5x performance and 9.5x energy efficiency improvements, compared to highly optimized parallel execution on the cores. Furthermore, we explore the requirements for end-to-end inference of a full mobile-grade DNN (MobileNetV2) in terms of IMC array resources, by scaling up our heterogeneous architecture to a multi-array accelerator. Our results show that our solution, on the end-to-end inference of the MobileNetV2, is one order of magnitude better in terms of execution latency than existing programmable architectures and two orders of magnitude better than state-of-the-art heterogeneous solutions integrating in-memory computing analog cores., Comment: 14 pages (not including final biography page), 13 figures (excluded authors pictures)

Deep Learning, and in particular, Deep Neural Network (DNN) is nowadays widely used in many scenarios, including safety-critical applications such as autonomous driving. In this context, besides energy efficiency and performance, reliability plays a crucial role since a system failure can jeopardize human life. As with any other device, the reliability of hardware architectures running DNNs has to be evaluated, usually through costly fault injection campaigns. This paper explores the approximation and fault resiliency of DNN accelerators. We propose to use approximate (AxC) arithmetic circuits to agilely emulate errors in hardware without performing fault injection on the DNN. To allow fast evaluation of AxC DNN, we developed an efficient GPU-based simulation framework. Further, we propose a fine-grain analysis of fault resiliency by examining fault propagation and masking in networks, 10 pages, 6 tables, 9 figures

Nowadays, the extensive exploitation of Deep Neural Networks (DNNs) in safety-critical applications raises new reliability concerns. In practice, methods for fault injection by emulation in hardware are efficient and widely used to study the resilience of DNN architectures for mitigating reliability issues already at the early design stages. However, the state-of-the-art methods for fault injection by emulation incur a spectrum of time-, design- and control-complexity problems. To overcome these issues, a novel resiliency assessment method called APPRAISER is proposed that applies functional approximation for a non-conventional purpose and employs approximate computing errors for its interest. By adopting this concept in the resiliency assessment domain, APPRAISER provides thousands of times speed-up in the assessment process, while keeping high accuracy of the analysis. In this paper, APPRAISER is validated by comparing it with state-of-the-art approaches for fault injection by emulation in FPGA. By this, the feasibility of the idea is demonstrated, and a new perspective in resiliency evaluation for DNNs is opened., 5 pages, 2 tables, 6 figures

Surveillance systems, autonomous vehicles, human monitoring systems, and video retrieval are just few of the many applications in which 3D Convolutional Neural Networks are exploited. However, their extensive use is restricted by their high computational and memory requirements, especially when integrated into systems with limited resources. This study proposes a toolflow that optimises the mapping of 3D CNN models for Human Action Recognition onto FPGA devices, taking into account FPGA resources and off-chip memory characteristics. The proposed system employs Synchronous Dataflow (SDF) graphs to model the designs and introduces transformations to expand and explore the design space, resulting in high-throughput designs. A variety of 3D CNN models were evaluated using the proposed toolflow on multiple FPGA devices, demonstrating its potential to deliver competitive performance compared to earlier hand-tuned and model-specific designs., 7 pages, 3 figures, 4 tables. arXiv admin note: substantial text overlap with arXiv:2305.18479

Passive resistive random access memory (ReRAM) crossbar arrays, a promising emerging technology used for analog matrix-vector multiplications, are far superior to their active (1T1R) counterparts in terms of the integration density. However, current transfers of neural network weights into the conductance state of the memory devices in the crossbar architecture are accompanied by significant losses in precision due to hardware variabilities such as sneak path currents, biasing scheme effects and conductance tuning imprecision. In this work, training approaches that adapt techniques such as dropout, the reparametrization trick and regularization to TiO2 crossbar variabilities are proposed in order to generate models that are better adapted to their hardware transfers. The viability of this approach is demonstrated by comparing the outputs and precision of the proposed hardware-aware network with those of a regular fully connected network over a few thousand weight transfers using the half moons dataset in a simulation based on experimental data. For the neural network trained using the proposed hardware-aware method, 79.5% of the test set's data points can be classified with an accuracy of 95% or higher, while only 18.5% of the test set's data points can be classified with this accuracy by the regularly trained neural network., 15 pages, 11 figures

Conventional multiply-accumulate (MAC) operations have long dominated computation time for deep neural networks (DNNs). Recently, product quantization (PQ) has been successfully applied to these workloads, replacing MACs with memory lookups to pre-computed dot products. While this property makes PQ an attractive solution for model acceleration, little is understood about the associated trade-offs in terms of compute and memory footprint, and the impact on accuracy. Our empirical study investigates the impact of different PQ settings and training methods on layerwise reconstruction error and end-to-end model accuracy. When studying the efficiency of deploying PQ DNNs, we find that metrics such as FLOPs, number of parameters, and even CPU/GPU performance, can be misleading. To address this issue, and to more fairly assess PQ in terms of hardware efficiency, we design the first custom hardware accelerator to evaluate the speed and efficiency of running PQ models. We identify PQ configurations that are able to improve performance-per-area for ResNet20 by 40%-104%, even when compared to a highly optimized conventional DNN accelerator. Our hardware performance outperforms recent PQ solutions by 4x, with only a 0.6% accuracy degradation. This work demonstrates the practical and hardware-aware design of PQ models, paving the way for wider adoption of this emerging DNN approximation methodology.

In this study, we introduce NeuralMatrix, a novel framework that enables the computation of versatile deep neural networks (DNNs) on a single general matrix multiplication (GEMM) accelerator. The proposed approach overcomes the specificity limitations of ASIC-based accelerators while achieving application-specific acceleration levels compared to general-purpose processors such as CPUs and GPUs. We address the challenges of mapping both linear and nonlinear operations in DNN computation to general matrix multiplications and the impact of using a GEMM accelerator on DNN inference accuracy. Extensive experiments are conducted on various DNN models from three popular categories (i.e., CNN, Transformers, and GNN) as illustrative backbone models. Our results demonstrate that DNNs suffer only up to a 2.02% accuracy loss after being converted to general matrix multiplication, while achieving 113x to 19.44x improvements in throughput per power compared to CPUs and GPUs., 12 pages, 4 figures, Submitted to 37th Conference on Neural Information Processing Systems (NeurIPS 2023)

Compute-in-Memory (CiM) utilizing non-volatile memory (NVM) devices presents a highly promising and efficient approach for accelerating deep neural networks (DNNs). By concurrently storing network weights and performing matrix operations within the same crossbar structure, CiM accelerators offer DNN inference acceleration with minimal area requirements and exceptional energy efficiency. However, the stochasticity and intrinsic variations of NVM devices often lead to performance degradation, such as reduced classification accuracy, compared to expected outcomes. Although several methods have been proposed to mitigate device variation and enhance robustness, most of them rely on overall modulation and lack constraints on the training process. Drawing inspiration from the negative feedback mechanism, we introduce a novel training approach that uses a multi-exit mechanism as negative feedback to enhance the performance of DNN models in the presence of device variation. Our negative feedback training method surpasses state-of-the-art techniques by achieving an impressive improvement of up to 12.49% in addressing DNN robustness against device variation.

Logic synthesis is the first and most vital step in chip design. This steps converts a chip specification written in a hardware description language (such as Verilog) into an optimized implementation using Boolean logic gates. State-of-the-art logic synthesis algorithms have a large number of logic minimization heuristics, typically applied sequentially based on human experience and intuition. The choice of the order greatly impacts the quality (e.g., area and delay) of the synthesized circuit. In this paper, we propose INVICTUS, a model-based offline reinforcement learning (RL) solution that automatically generates a sequence of logic minimization heuristics ("synthesis recipe") based on a training dataset of previously seen designs. A key challenge is that new designs can range from being very similar to past designs (e.g., adders and multipliers) to being completely novel (e.g., new processor instructions). %Compared to prior work, INVICTUS is the first solution that uses a mix of RL and search methods joint with an online out-of-distribution detector to generate synthesis recipes over a wide range of benchmarks. Our results demonstrate significant improvement in area-delay product (ADP) of synthesized circuits with up to 30\% improvement over state-of-the-art techniques. Moreover, INVICTUS achieves up to $6.3\times$ runtime reduction (iso-ADP) compared to the state-of-the-art., 20 pages, 8 figures and 15 tables

This paper introduces a novel optimization framework for deep neural network (DNN) hardware accelerators, enabling the rapid development of customized and automated design flows. More specifically, our approach aims to automate the selection and configuration of low-level optimization techniques, encompassing DNN and FPGA low-level optimizations. We introduce novel optimization and transformation tasks for building design-flow architectures, which are highly customizable and flexible, thereby enhancing the performance and efficiency of DNN accelerators. Our results demonstrate considerable reductions of up to 92\% in DSP usage and 89\% in LUT usage for two networks, while maintaining accuracy and eliminating the need for human effort or domain expertise. In comparison to state-of-the-art approaches, our design achieves higher accuracy and utilizes three times fewer DSP resources, underscoring the advantages of our proposed framework., Comment: 5 pages, Accepted at FPL'23

Due to the unavailability of routing information in design stages prior to detailed routing (DR), the tasks of timing prediction and optimization pose major challenges. Inaccurate timing prediction wastes design effort, hurts circuit performance, and may lead to design failure. This work focuses on timing prediction after clock tree synthesis and placement legalization, which is the earliest opportunity to time and optimize a "complete" netlist. The paper first documents that having "oracle knowledge" of the final post-DR parasitics enables post-global routing (GR) optimization to produce improved final timing outcomes. To bridge the gap between GR-based parasitic and timing estimation and post-DR results during post-GR optimization, machine learning (ML)-based models are proposed, including the use of features for macro blockages for accurate predictions for designs with macros. Based on a set of experimental evaluations, it is demonstrated that these models show higher accuracy than GR-based timing estimation. When used during post-GR optimization, the ML-based models show demonstrable improvements in post-DR circuit performance. The methodology is applied to two different tool flows - OpenROAD and a commercial tool flow - and results on 45nm bulk and 12nm FinFET enablements show improvements in post-DR slack metrics without increasing congestion. The models are demonstrated to be generalizable to designs generated under different clock period constraints and are robust to training data with small levels of noise.

Artificial Intelligence (AI) and, in particular, Machine Learning (ML) have emerged to be utilized in various applications due to their capability to learn how to solve complex problems. Over the last decade, rapid advances in ML have presented Deep Neural Networks (DNNs) consisting of a large number of neurons and layers. DNN Hardware Accelerators (DHAs) are leveraged to deploy DNNs in the target applications. Safety-critical applications, where hardware faults/errors would result in catastrophic consequences, also benefit from DHAs. Therefore, the reliability of DNNs is an essential subject of research. In recent years, several studies have been published accordingly to assess the reliability of DNNs. In this regard, various reliability assessment methods have been proposed on a variety of platforms and applications. Hence, there is a need to summarize the state of the art to identify the gaps in the study of the reliability of DNNs. In this work, we conduct a Systematic Literature Review (SLR) on the reliability assessment methods of DNNs to collect relevant research works as much as possible, present a categorization of them, and address the open challenges. Through this SLR, three kinds of methods for reliability assessment of DNNs are identified including Fault Injection (FI), Analytical, and Hybrid methods. Since the majority of works assess the DNN reliability by FI, we characterize different approaches and platforms of the FI method comprehensively. Moreover, Analytical and Hybrid methods are propounded. Thus, different reliability assessment methods for DNNs have been elaborated on their conducted DNN platforms and reliability evaluation metrics. Finally, we highlight the advantages and disadvantages of the identified methods and address the open challenges in the research area., 42 pages, 15 figures, 3 tables, 201 references. The paper has been reviewed and revised 2 times and is under the 3rd review

The emergence of the Internet of Things (IoT) has resulted in a remarkable amount of data generated on edge devices, which are often processed using AI algorithms. On-device learning enables edge platforms to continually adapt the AI models to user personal data and further allows for a better service quality. However, AI training on resource-limited devices is extremely difficult because of the intensive computing workload and the significant amount of on-chip memory consumption exacted by deep neural networks (DNNs). To mitigate this, we propose to use embedded dynamic random-access memory (eDRAM) as the main storage medium of training data. Compared with static random-access memory (SRAM), eDRAM introduces more than $2\times$ improvement on storage density, enabling reduced off-chip memory traffic. However, to keep the stored data intact, eDRAM is required to perform the power-hungry data refresh operations. eDRAM refresh can be eliminated if the data is stored for a period of time that is shorter than the eDRAM retention time. To achieve this, we design a novel reversible DNN architecture that enables a significantly reduced data lifetime during the training process and removes the need for eDRAM refresh. We further design an efficient on-device training engine, termed~\textit{CAMEL}, that uses eDRAM as the main on-chip memory. CAMEL enables the intermediate results during training to fit fully in on-chip eDRAM arrays and completely eliminates the off-chip DRAM traffic during the training process. We evaluate our CAMEL system on multiple DNNs with different datasets, demonstrating a more than $3\times$ saving on total DNN training energy consumption than the other baselines, while achieving a similar (even better) performance in validation accuracy.

Hardware Trojans (HTs) are undesired design or manufacturing modifications that can severely alter the security and functionality of digital integrated circuits. HTs can be inserted according to various design criteria, e.g., nets switching activity, observability, controllability, etc. However, to our knowledge, most HT detection methods are only based on a single criterion, i.e., nets switching activity. This paper proposes a multi-criteria reinforcement learning (RL) HT detection tool that features a tunable reward function for different HT detection scenarios. The tool allows for exploring existing detection strategies and can adapt new detection scenarios with minimal effort. We also propose a generic methodology for comparing HT detection methods fairly. Our preliminary results show an average of 84.2% successful HT detection in ISCAS-85 benchmark

Fault-Aware Training (FAT) has emerged as a highly effective technique for addressing permanent faults in DNN accelerators, as it offers fault mitigation without significant performance or accuracy loss, specifically at low and moderate fault rates. However, it leads to very high retraining overheads, especially when used for large DNNs designed for complex AI applications. Moreover, as each fabricated chip can have a distinct fault pattern, FAT is required to be performed for each faulty chip individually, considering its unique fault map, which further aggravates the problem. To reduce the overheads of FAT while maintaining its benefits, we propose (1) the concepts of resilience-driven retraining amount selection, and (2) resilience-driven grouping and fusion of multiple fault maps (belonging to different chips) to perform consolidated retraining for a group of faulty chips. To realize these concepts, in this work, we present a novel framework, eFAT, that computes the resilience of a given DNN to faults at different fault rates and with different levels of retraining, and it uses that knowledge to build a resilience map given a user-defined accuracy constraint. Then, it uses the resilience map to compute the amount of retraining required for each chip, considering its unique fault map. Afterward, it performs resilience and reward-driven grouping and fusion of fault maps to further reduce the number of retraining iterations required for tuning the given DNN for the given set of faulty chips. We demonstrate the effectiveness of our framework for a systolic array-based DNN accelerator experiencing permanent faults in the computational array. Our extensive results for numerous chips show that the proposed technique significantly reduces the retraining cost when used for tuning a DNN for multiple faulty chips., 8 pages, 13 figures

In this paper, we present microring resonator (MRR) based polymorphic E-O circuits and architectures that can be employed for high-speed and energy-efficient non-binary reconfigurable computing. Our polymorphic E-O circuits can be dynamically programmed to implement different logic and arithmetic functions at different times. They can provide compactness and polymorphism to consequently improve operand handling, reduce idle time, and increase amortization of area and static power overheads. When combined with flexible photodetectors with the innate ability to accumulate a high number of optical pulses in situ, our circuits can support energy-efficient processing of data in non-binary formats such as stochastic/unary and high-dimensional reservoir formats. Furthermore, our polymorphic E-O circuits enable configurable E-O computing accelerator architectures for processing binarized and integer quantized convolutional neural networks (CNNs). We compare our designed polymorphic E-O circuits and architectures to several circuits and architectures from prior works in terms of area, latency, and energy consumption.

We develop an end-to-end workflow for the training and implementation of co-designed neural networks (NNs) for efficient field-programmable gate array (FPGA) and application-specific integrated circuit (ASIC) hardware. Our approach leverages Hessian-aware quantization (HAWQ) of NNs, the Quantized Open Neural Network Exchange (QONNX) intermediate representation, and the hls4ml tool flow for transpiling NNs into FPGA and ASIC firmware. This makes efficient NN implementations in hardware accessible to nonexperts, in a single open-sourced workflow that can be deployed for real-time machine learning applications in a wide range of scientific and industrial settings. We demonstrate the workflow in a particle physics application involving trigger decisions that must operate at the 40 MHz collision rate of the CERN Large Hadron Collider (LHC). Given the high collision rate, all data processing must be implemented on custom ASIC and FPGA hardware within a strict area and latency. Based on these constraints, we implement an optimized mixed-precision NN classifier for high-momentum particle jets in simulated LHC proton-proton collisions., 19 pages, 6 figures, 2 tables

To maximize the performance and energy efficiency of Spiking Neural Network (SNN) processing on resource-constrained embedded systems, specialized hardware accelerators/chips are employed. However, these SNN chips may suffer from permanent faults which can affect the functionality of weight memory and neuron behavior, thereby causing potentially significant accuracy degradation and system malfunctioning. Such permanent faults may come from manufacturing defects during the fabrication process, and/or from device/transistor damages (e.g., due to wear out) during the run-time operation. However, the impact of permanent faults in SNN chips and the respective mitigation techniques have not been thoroughly investigated yet. Toward this, we propose RescueSNN, a novel methodology to mitigate permanent faults in the compute engine of SNN chips without requiring additional retraining, thereby significantly cutting down the design time and retraining costs, while maintaining the throughput and quality. The key ideas of our RescueSNN methodology are (1) analyzing the characteristics of SNN under permanent faults; (2) leveraging this analysis to improve the SNN fault-tolerance through effective fault-aware mapping (FAM); and (3) devising lightweight hardware enhancements to support FAM. Our FAM technique leverages the fault map of SNN compute engine for (i) minimizing weight corruption when mapping weight bits on the faulty memory cells, and (ii) selectively employing faulty neurons that do not cause significant accuracy degradation to maintain accuracy and throughput, while considering the SNN operations and processing dataflow. The experimental results show that our RescueSNN improves accuracy by up to 80% while maintaining the throughput reduction below 25% in high fault rate (e.g., 0.5 of the potential fault locations), as compared to running SNNs on the faulty chip without mitigation., Accepted for publication at Frontiers in Neuroscience - Section Neuromorphic Engineering

We present the MEMA framework for the easy and quick derivation of efficient inference runtimes that minimize external memory accesses for matrix multiplication on TinyML systems. The framework accounts for hardware resource constraints and problem sizes in analytically determining optimized schedules and kernels that minimize memory accesses. MEMA provides a solution to a well-known problem in the current practice, that is, optimal schedules tend to be found only through a time consuming and heuristic search of a large scheduling space. We compare the performance of runtimes derived from MEMA to existing state-of-the-art libraries on ARM-based TinyML systems. For example, for neural network benchmarks on the ARM Cortex-M4, we achieve up to a 1.8x speedup and 44% energy reduction over CMSIS-NN., Accepted as a full paper by the TinyML Research Symposium 2023

Spiking Neural Networks (SNNs) have shown capabilities of achieving high accuracy under unsupervised settings and low operational power/energy due to their bio-plausible computations. Previous studies identified that DRAM-based off-chip memory accesses dominate the energy consumption of SNN processing. However, state-of-the-art works do not optimize the DRAM energy-per-access, thereby hindering the SNN-based systems from achieving further energy efficiency gains. To substantially reduce the DRAM energy-per-access, an effective solution is to decrease the DRAM supply voltage, but it may lead to errors in DRAM cells (i.e., so-called approximate DRAM). Towards this, we propose \textit{EnforceSNN}, a novel design framework that provides a solution for resilient and energy-efficient SNN inference using reduced-voltage DRAM for embedded systems. The key mechanisms of our EnforceSNN are: (1) employing quantized weights to reduce the DRAM access energy; (2) devising an efficient DRAM mapping policy to minimize the DRAM energy-per-access; (3) analyzing the SNN error tolerance to understand its accuracy profile considering different bit error rate (BER) values; (4) leveraging the information for developing an efficient fault-aware training (FAT) that considers different BER values and bit error locations in DRAM to improve the SNN error tolerance; and (5) developing an algorithm to select the SNN model that offers good trade-offs among accuracy, memory, and energy consumption. The experimental results show that our EnforceSNN maintains the accuracy (i.e., no accuracy loss for BER less-or-equal 10^-3) as compared to the baseline SNN with accurate DRAM, while achieving up to 84.9\% of DRAM energy saving and up to 4.1x speed-up of DRAM data throughput across different network sizes., Accepted for publication at Frontiers in Neuroscience - Section Neuromorphic Engineering

As deep learning advances, edge devices and lightweight neural networks are becoming more important. To reduce latency in the AI accelerator, it's essential to not only reduce FLOPs but also enhance hardware performance. We proposed an arithmetic intensity balancing convolution (ABConv) to address the issue of the overall intensity being limited by the small weight arithmetic intensity for convolution with a small spatial size. ABConv increased the maximum bound of overall arithmetic intensity and significantly reduced latency, without sacrificing accuracy. We tested the latency and hardware performance of ABConv on the Arm Ethos-U65 NPU in various configurations and used it to replace some of MobileNetV1 and ResNet50 in image classification for CIFAR100., Accepted paper at the On-Device Intelligence Workshop in conjunction with MLSys Conference 2023

Processing-in-memory (PIM) promises to alleviate the data movement bottleneck in modern computing systems. However, current real-world PIM systems have the inherent disadvantage that their hardware is more constrained than in conventional processors (CPU, GPU), due to the difficulty and cost of building processing elements near or inside the memory. As a result, general-purpose PIM architectures support fairly limited instruction sets and struggle to execute complex operations such as transcendental functions and other hard-to-calculate operations (e.g., square root). These operations are particularly important for some modern workloads, e.g., activation functions in machine learning applications. In order to provide support for transcendental (and other hard-to-calculate) functions in general-purpose PIM systems, we present \emph{TransPimLib}, a library that provides CORDIC-based and LUT-based methods for trigonometric functions, hyperbolic functions, exponentiation, logarithm, square root, etc. We develop an implementation of TransPimLib for the UPMEM PIM architecture and perform a thorough evaluation of TransPimLib's methods in terms of performance and accuracy, using microbenchmarks and three full workloads (Blackscholes, Sigmoid, Softmax). We open-source all our code and datasets at~\url{https://github.com/CMU-SAFARI/transpimlib}., Our open-source software is available at https://github.com/CMU-SAFARI/transpimlib

This paper proposes a special-purpose system to achieve high-accuracy and high-efficiency machine learning (ML) molecular dynamics (MD) calculations. The system consists of field programmable gate array (FPGA) and application specific integrated circuit (ASIC) working in heterogeneous parallelization. To be specific, a multiplication-less neural network (NN) is deployed on the non-von Neumann (NvN)-based ASIC (SilTerra 180 nm process) to evaluate atomic forces, which is the most computationally expensive part of MD. All other calculations of MD are done using FPGA (Xilinx XC7Z100). It is shown that, to achieve similar-level accuracy, the proposed NvN-based system based on low-end fabrication technologies (180 nm) is 1.6x faster and 10^2-10^3x more energy efficiency than state-of-the-art vN based MLMD using graphics processing units (GPUs) based on much more advanced technologies (12 nm), indicating superiority of the proposed NvN-based heterogeneous parallel architecture.