Search

Rapidly evolving artificial intelligence and machine learning applications require ever-increasing computational capabilities, while monolithic 2D design technologies approach their limits. Heterogeneous integration of smaller chiplets using a 2.5D silicon interposer and 3D packaging has emerged as a promising paradigm to address this limit and meet performance demands. These approaches offer a significant cost reduction and higher manufacturing yield than monolithic 2D integrated circuits. However, the compact arrangement and high compute density exacerbate the thermal management challenges, potentially compromising performance. Addressing these thermal modeling challenges is critical, especially as system sizes grow and different design stages require varying levels of accuracy and speed. Since no single thermal modeling technique meets all these needs, this paper introduces MFIT, a range of multi-fidelity thermal models that effectively balance accuracy and speed. These multi-fidelity models can enable efficient design space exploration and runtime thermal management. Our extensive testing on systems with 16, 36, and 64 2.5D integrated chiplets and 16x3 3D integrated chiplets demonstrates that these models can reduce execution times from days to mere seconds and milliseconds with negligible loss in accuracy., Comment: Preprint for MFIT: Multi-Fidelity Thermal Modeling for 2.5D and 3D Multi-Chiplet Architectures

Transformers have revolutionized deep learning and generative modeling to enable unprecedented advancements in natural language processing tasks and beyond. However, designing hardware accelerators for executing transformer models is challenging due to the wide variety of computing kernels involved in the transformer architecture. Existing accelerators are either inadequate to accelerate end-to-end transformer models or suffer notable thermal limitations. In this paper, we propose the design of a three-dimensional heterogeneous architecture referred to as HeTraX specifically optimized to accelerate end-to-end transformer models. HeTraX employs hardware resources aligned with the computational kernels of transformers and optimizes both performance and energy. Experimental results show that HeTraX outperforms existing state-of-the-art by up to 5.6x in speedup and improves EDP by 14.5x while ensuring thermally feasibility., Comment: Presented at ACM/IEEE International Symposium on Low Power Electronics and Design (ISLPED-24)

Traditional digital implementations of neural accelerators are limited by high power and area overheads, while analog and non-CMOS implementations suffer from noise, device mismatch, and reliability issues. This paper introduces a CMOS Look-Up Table (LUT)-based Neural Accelerator (LUT-NA) framework that reduces the power, latency, and area consumption of traditional digital accelerators through pre-computed, faster look-ups while avoiding noise and mismatch of analog circuits. To solve the scalability issues of conventional LUT-based computation, we split the high-precision multiply and accumulate (MAC) operations into lower-precision MACs using a divide-and-conquer-based approach. We show that LUT-NA achieves up to $29.54\times$ lower area with $3.34\times$ lower energy per inference task than traditional LUT-based techniques and up to $1.23\times$ lower area with $1.80\times$ lower energy per inference task than conventional digital MAC-based techniques (Wallace Tree/Array Multipliers) without retraining and without affecting accuracy, even on lottery ticket pruned (LTP) models that already reduce the number of required MAC operations by up to 98%. Finally, we introduce mixed precision analysis in LUT-NA framework for various LTP models (VGG11, VGG19, Resnet18, Resnet34, GoogleNet) that achieved up to $32.22\times$-$50.95\times$ lower area across models with $3.68\times$-$6.25\times$ lower energy per inference than traditional LUT-based techniques, and up to $1.35\times$-$2.14\times$ lower area requirement with $1.99\times$-$3.38\times$ lower energy per inference across models as compared to conventional digital MAC-based techniques with $\sim$1% accuracy loss., Comment: 7 pages

Processing-in-memory (PIM) has emerged as an enabler for the energy-efficient and high-performance acceleration of deep learning (DL) workloads. Resistive random-access memory (ReRAM) is one of the most promising technologies to implement PIM. However, as the complexity of Deep convolutional neural networks (DNNs) grows, we need to design a manycore architecture with multiple ReRAM-based processing elements (PEs) on a single chip. Existing PIM-based architectures mostly focus on computation while ignoring the role of communication. ReRAM-based tiled manycore architectures often involve many Processing Elements (PEs), which need to be interconnected via an efficient on-chip communication infrastructure. Simply allocating more resources (ReRAMs) to speed up only computation is ineffective if the communication infrastructure cannot keep up with it. In this paper, we highlight the design principles of a dataflow-aware PIM-enabled manycore platform tailor-made for various types of DL workloads. We consider the design challenges with both 2.5D interposer- and 3D integration-enabled architectures., Comment: Presented at DATE Conference, Valencia, Spain 2024

Resistive random-access memory (ReRAM)-based processing-in-memory (PIM) architecture is an attractive solution for training Graph Neural Networks (GNNs) on edge platforms. However, the immature fabrication process and limited write endurance of ReRAMs make them prone to hardware faults, thereby limiting their widespread adoption for GNN training. Further, the existing fault-tolerant solutions prove inadequate for effectively training GNNs in the presence of faults. In this paper, we propose a fault-aware framework referred to as FARe that mitigates the effect of faults during GNN training. FARe outperforms existing approaches in terms of both accuracy and timing overhead. Experimental results demonstrate that FARe framework can restore GNN test accuracy by 47.6% on faulty ReRAM hardware with a ~1% timing overhead compared to the fault-free counterpart., Comment: This paper has been accepted to the conference DATE (Design, Automation and Test in Europe) - 2024

Transformers have revolutionized deep learning and generative modeling, enabling unprecedented advancements in natural language processing tasks. However, the size of transformer models is increasing continuously, driven by enhanced capabilities across various deep-learning tasks. This trend of ever-increasing model size has given rise to new challenges in terms of memory and computing requirements. Conventional computing platforms, including GPUs, suffer from suboptimal performance due to the memory demands imposed by models with millions/billions of parameters. The emerging chiplet-based platforms provide a new avenue for compute- and data-intensive machine learning (ML) applications enabled by a Network-on-Interposer (NoI). However, designing suitable hardware accelerators for executing Transformer inference workloads is challenging due to a wide variety of complex computing kernels in the Transformer architecture. In this paper, we leverage chiplet-based heterogeneous integration (HI) to design a high-performance and energy-efficient multi-chiplet platform to accelerate transformer workloads. We demonstrate that the proposed NoI architecture caters to the data access patterns inherent in a transformer model. The optimized placement of the chiplets and the associated NoI links and routers enable superior performance compared to the state-of-the-art hardware accelerators. The proposed NoI-based architecture demonstrates scalability across varying transformer models and improves latency and energy efficiency by up to 22.8x and 5.36x respectively., Comment: Preprint for a Heterogeneous Chiplet Architecture for Accelerating End-to-End Transformer Models

Resistive random access memory (ReRAM)-based processing-in-memory (PIM) architectures have demonstrated great potential to accelerate Deep Neural Network (DNN) training/inference. However, the computational accuracy of analog PIM is compromised due to the non-idealities, such as the conductance variation of ReRAM cells. The impact of these non-idealities worsens as the number of concurrently activated wordlines and bitlines increases. To guarantee computational accuracy, only a limited number of wordlines and bitlines of the crossbar array can be turned on concurrently, significantly reducing the achievable parallelism of the architecture. While the constraints on parallelism limit the efficiency of the accelerators, they also provide a new opportunity for fine-grained mixed-precision quantization. To enable efficient DNN inference on practical ReRAM-based accelerators, we propose an algorithm-architecture co-design framework called \underline{B}lock-\underline{W}ise mixed-precision \underline{Q}uantization (BWQ). At the algorithm level, BWQ-A introduces a mixed-precision quantization scheme at the block level, which achieves a high weight and activation compression ratio with negligible accuracy degradation. We also present the hardware architecture design BWQ-H, which leverages the low-bit-width models achieved by BWQ-A to perform high-efficiency DNN inference on ReRAM devices. BWQ-H also adopts a novel precision-aware weight mapping method to increase the ReRAM crossbar's throughput. Our evaluation demonstrates the effectiveness of BWQ, which achieves a 6.08x speedup and a 17.47x energy saving on average compared to existing ReRAM-based architectures., Comment: 12 pages, 13 figures

Training machine learning (ML) models at the edge (on-chip training on end user devices) can address many pressing challenges including data privacy/security, increase the accessibility of ML applications to different parts of the world by reducing the dependence on the communication fabric and the cloud infrastructure, and meet the real-time requirements of AR/VR applications. However, existing edge platforms do not have sufficient computing capabilities to support complex ML tasks such as training large CNNs. ReRAM-based architectures offer high-performance yet energy efficient computing platforms for on-chip CNN training/inferencing. However, ReRAM-based architectures are not scalable with the size of the CNN. Larger CNNs have more weights, which requires more ReRAM cells that cannot be integrated in a single chip. Moreover, training larger CNNs on-chip will require higher power, which cannot be afforded by these smaller devices. Pruning is an effective way to solve this problem. However, existing pruning techniques are either targeted for inferencing only, or they are not crossbar-aware. This leads to sub-optimal hardware savings and performance benefits for CNN training on ReRAM-based architectures. In this paper, we address this problem by proposing a novel crossbar-aware pruning strategy, referred as ReaLPrune, which can prune more than 90% of CNN weights. The pruned model can be trained from scratch without any accuracy loss. Experimental results indicate that ReaLPrune reduces hardware requirements by 77.2% and accelerates CNN training by ~20X compared to unpruned CNNs. ReaLPrune also outperforms other crossbar-aware pruning techniques in terms of both performance and hardware savings. In addition, ReaLPrune is equally effective for diverse datasets and more complex CNNs, Comment: 13 pages, 9 figures

Resistive random-access memory (ReRAM) is a promising technology for designing hardware accelerators for deep neural network (DNN) inferencing. However, stochastic noise in ReRAM crossbars can degrade the DNN inferencing accuracy. We propose the design and optimization of a high-performance, area-and energy-efficient ReRAM-based hardware accelerator to achieve robust DNN inferencing in the presence of stochastic noise. We make two key technical contributions. First, we propose a stochastic-noise-aware training method, referred to as ReSNA, to improve the accuracy of DNN inferencing on ReRAM crossbars with stochastic noise. Second, we propose an information-theoretic algorithm, referred to as CF-MESMO, to identify the Pareto set of solutions to trade-off multiple objectives, including inferencing accuracy, area overhead, execution time, and energy consumption. The main challenge in this context is that executing the ReSNA method to evaluate each candidate ReRAM design is prohibitive. To address this challenge, we utilize the continuous-fidelity evaluation of ReRAM designs associated with prohibitive high computation cost by varying the number of training epochs to trade-off accuracy and cost. CF-MESMO iteratively selects the candidate ReRAM design and fidelity pair that maximizes the information gained per unit computation cost about the optimal Pareto front. Our experiments on benchmark DNNs show that the proposed algorithms efficiently uncover high-quality Pareto fronts. On average, ReSNA achieves 2.57% inferencing accuracy improvement for ResNet20 on the CIFAR-10 dataset with respect to the baseline configuration. Moreover, CF-MESMO algorithm achieves 90.91% reduction in computation cost compared to the popular multi-objective optimization algorithm NSGA-II to reach the best solution from NSGA-II., Comment: To appear in ICCAD 2021

Mobile system-on-chips (SoCs) are growing in their complexity and heterogeneity (e.g., Arm's Big-Little architecture) to meet the needs of emerging applications, including games and artificial intelligence. This makes it very challenging to optimally manage the resources (e.g., controlling the number and frequency of different types of cores) at runtime to meet the desired trade-offs among multiple objectives such as performance and energy. This paper proposes a novel information-theoretic framework referred to as PaRMIS to create Pareto-optimal resource management policies for given target applications and design objectives. PaRMIS specifies parametric policies to manage resources and learns statistical models from candidate policy evaluation data in the form of target design objective values. The key idea is to select a candidate policy for evaluation in each iteration guided by statistical models that maximize the information gain about the true Pareto front. Experiments on a commercial heterogeneous SoC show that PaRMIS achieves better Pareto fronts and is easily usable to optimize complex objectives (e.g., performance per Watt) when compared to prior methods., Comment: To be published in proceedings DAC

We consider the task of photo-realistic unconditional image generation (generate high quality, diverse samples that carry the same visual content as the image) on mobile platforms using Generative Adversarial Networks (GANs). In this paper, we propose a novel approach to trade-off image generation accuracy of a GAN for the energy consumed (compute) at run-time called Scale-Energy Tradeoff GAN (SETGAN). GANs usually take a long time to train and consume a huge memory hence making it difficult to run on edge devices. The key idea behind SETGAN for an image generation task is for a given input image, we train a GAN on a remote server and use the trained model on edge devices. We use SinGAN, a single image unconditional generative model, that contains a pyramid of fully convolutional GANs, each responsible for learning the patch distribution at a different scale of the image. During the training process, we determine the optimal number of scales for a given input image and the energy constraint from the target edge device. Results show that with SETGAN's unique client-server-based architecture, we were able to achieve a 56% gain in energy for a loss of 3% to 12% SSIM accuracy. Also, with the parallel multi-scale training, we obtain around 4x gain in training time on the server.

Graph Neural Network (GNN) is a variant of Deep Neural Networks (DNNs) operating on graphs. However, GNNs are more complex compared to traditional DNNs as they simultaneously exhibit features of both DNN and graph applications. As a result, architectures specifically optimized for either DNNs or graph applications are not suited for GNN training. In this work, we propose a 3D heterogeneous manycore architecture for on-chip GNN training to address this problem. The proposed architecture, ReGraphX, involves heterogeneous ReRAM crossbars to fulfill the disparate requirements of both DNN and graph computations simultaneously. The ReRAM-based architecture is complemented with a multicast-enabled 3D NoC to improve the overall achievable performance. We demonstrate that ReGraphX outperforms conventional GPUs by up to 3.5X (on an average 3X) in terms of execution time, while reducing energy consumption by as much as 11X., Comment: This paper has been accepted and presented at Design Automation and Test in Europe (DATE) 2021

Heterogeneous manycore architectures are the key to efficiently execute compute- and data-intensive applications. Through silicon via (TSV)-based 3D manycore system is a promising solution in this direction as it enables integration of disparate computing cores on a single system. However, the achievable performance of conventional through-silicon-via (TSV)-based 3D systems is ultimately bottlenecked by the horizontal wires (wires in each planar die). Moreover, current TSV 3D architectures suffer from thermal limitations. Hence, TSV-based architectures do not realize the full potential of 3D integration. Monolithic 3D (M3D) integration, a breakthrough technology to achieve - More Moore and More Than Moore - and opens up the possibility of designing cores and associated network routers using multiple layers by utilizing monolithic inter-tier vias (MIVs) and hence, reducing the effective wire length. Compared to TSV-based 3D ICs, M3D offers the true benefits of vertical dimension for system integration: the size of a MIV used in M3D is over 100x smaller than a TSV. In this work, we demonstrate how M3D-enabled vertical core and uncore elements offer significant performance and thermal improvements in manycore heterogeneous architectures compared to its TSV-based counterpart. To overcome the difficult optimization challenges due to the large design space and complex interactions among the heterogeneous components (CPU, GPU, Last Level Cache, etc.) in an M3D-based manycore chip, we leverage novel design-space exploration algorithms to trade-off different objectives. The proposed M3D-enabled heterogeneous architecture, called HeM3D, outperforms its state-of-the-art TSV-equivalent counterpart by up to 18.3% in execution time while being up to 19 degrees Celcius cooler., Comment: This work has been accepted in ACM Transactions on Design Automation of Electronic Systems

Mobile platforms must satisfy the contradictory requirements of fast response time and minimum energy consumption as a function of dynamically changing applications. To address this need, system-on-chips (SoC) that are at the heart of these devices provide a variety of control knobs, such as the number of active cores and their voltage/frequency levels. Controlling these knobs optimally at runtime is challenging for two reasons. First, the large configuration space prohibits exhaustive solutions. Second, control policies designed offline are at best sub-optimal since many potential new applications are unknown at design-time. We address these challenges by proposing an online imitation learning approach. Our key idea is to construct an offline policy and adapt it online to new applications to optimize a given metric (e.g., energy). The proposed methodology leverages the supervision enabled by power-performance models learned at runtime. We demonstrate its effectiveness on a commercial mobile platform with 16 diverse benchmarks. Our approach successfully adapts the control policy to an unknown application after executing less than 25% of its instructions., Comment: This paper has been accepted to be published in a future issue of ACM TODAES

Monolithic 3D (M3D) technology enables high density integration, performance, and energy-efficiency by sequentially stacking tiers on top of each other. M3D-based network-on-chip (NoC) architectures can exploit these benefits by adopting tier partitioning for intra-router stages. However, conventional fabrication methods are infeasible for M3D-enabled designs due to temperature related issues. This has necessitated lower temperature and temperature-resilient techniques for M3D fabrication, leading to inferior performance of transistors in the top tier and interconnects in the bottom tier. The resulting inter-tier process variation leads to performance degradation of M3D-enabled NoCs. In this work, we demonstrate that without considering inter-tier process variation, an M3D-enabled NoC architecture overestimates the energy-delay-product (EDP) on average by 50.8% for a set of SPLASH-2 and PARSEC benchmarks. As a countermeasure, we adopt a process variation aware design approach. The proposed design and optimization method distribute the intra-router stages and inter-router links among the tiers to mitigate the adverse effects of process variation. Experimental results show that the NoC architecture under consideration improves the EDP by 27.4% on average across all benchmarks compared to the process-oblivious design., Comment: Submitted to IEEE TVLSI (Under Review)

Deep neural networks have seen tremendous success for different modalities of data including images, videos, and speech. This success has led to their deployment in mobile and embedded systems for real-time applications. However, making repeated inferences using deep networks on embedded systems poses significant challenges due to constrained resources (e.g., energy and computing power). To address these challenges, we develop a principled co-design approach. Building on prior work, we develop a formalism referred to as Coarse-to-Fine Networks (C2F Nets) that allow us to employ classifiers of varying complexity to make predictions. We propose a principled optimization algorithm to automatically configure C2F Nets for a specified trade-off between accuracy and energy consumption for inference. The key idea is to select a classifier on-the-fly whose complexity is proportional to the hardness of the input example: simple classifiers for easy inputs and complex classifiers for hard inputs. We perform comprehensive experimental evaluation using four different C2F Net architectures on multiple real-world image classification tasks. Our results show that optimized C2F Net can reduce the Energy Delay Product (EDP) by 27 to 60 percent with no loss in accuracy when compared to the baseline solution, where all predictions are made using the most complex classifier in C2F Net., Comment: Published in IEEE Trans. on CAD of Integrated Circuits and Systems

The rising use of deep learning and other big-data algorithms has led to an increasing demand for hardware platforms that are computationally powerful, yet energy-efficient. Due to the amount of data parallelism in these algorithms, high-performance 3D manycore platforms that incorporate both CPUs and GPUs present a promising direction. However, as systems use heterogeneity (e.g., a combination of CPUs, GPUs, and accelerators) to improve performance and efficiency, it becomes more pertinent to address the distinct and likely conflicting communication requirements (e.g., CPU memory access latency or GPU network throughput) that arise from such heterogeneity. Unfortunately, it is difficult to quickly explore the hardware design space and choose appropriate tradeoffs between these heterogeneous requirements. To address these challenges, we propose the design of a 3D Network-on-Chip (NoC) for heterogeneous manycore platforms that considers the appropriate design objectives for a 3D heterogeneous system and explores various tradeoffs using an efficient ML-based multi-objective optimization technique. The proposed design space exploration considers the various requirements of its heterogeneous components and generates a set of 3D NoC architectures that efficiently trades off these design objectives. Our findings show that by jointly considering these requirements (latency, throughput, temperature, and energy), we can achieve 9.6% better Energy-Delay Product on average at nearly iso-temperature conditions when compared to a thermally-optimized design for 3D heterogeneous NoCs. More importantly, our results suggest that our 3D NoCs optimized for a few applications can be generalized for unknown applications as well. Our results show that these generalized 3D NoCs only incur a 1.8% (36-tile system) and 1.1% (64-tile system) average performance loss compared to application-specific NoCs., Comment: Published in IEEE Transactions on Computers

Convolutional Neural Networks (CNNs) have shown a great deal of success in diverse application domains including computer vision, speech recognition, and natural language processing. However, as the size of datasets and the depth of neural network architectures continue to grow, it is imperative to design high-performance and energy-efficient computing hardware for training CNNs. In this paper, we consider the problem of designing specialized CPU-GPU based heterogeneous manycore systems for energy-efficient training of CNNs. It has already been shown that the typical on-chip communication infrastructures employed in conventional CPU-GPU based heterogeneous manycore platforms are unable to handle both CPU and GPU communication requirements efficiently. To address this issue, we first analyze the on-chip traffic patterns that arise from the computational processes associated with training two deep CNN architectures, namely, LeNet and CDBNet, to perform image classification. By leveraging this knowledge, we design a hybrid Network-on-Chip (NoC) architecture, which consists of both wireline and wireless links, to improve the performance of CPU-GPU based heterogeneous manycore platforms running the above-mentioned CNN training workloads. The proposed NoC achieves 1.8x reduction in network latency and improves the network throughput by a factor of 2.2 for training CNNs, when compared to a highly-optimized wireline mesh NoC. For the considered CNN workloads, these network-level improvements translate into 25% savings in full-system energy-delay-product (EDP). This demonstrates that the proposed hybrid NoC for heterogeneous manycore architectures is capable of significantly accelerating training of CNNs while remaining energy-efficient., Comment: Accepted in a future publication of IEEE Transactions on Computers

Tight collaboration between experts of machine learning and manycore system design is necessary to create a data-driven manycore design framework that integrates both learning and expert knowledge. Such a framework will be necessary to address the rising complexity of designing large-scale manycore systems and machine learning techniques., Comment: To appear in a future publication of IEEE Computer

A three-dimensional (3D) Network-on-Chip (NoC) enables the design of high performance and low power many-core chips. Existing 3D NoCs are inadequate for meeting the ever-increasing performance requirements of many-core processors since they are simple extensions of regular 2D architectures and they do not fully exploit the advantages provided by 3D integration. Moreover, the anticipated performance gain of a 3D NoC-enabled many-core chip may be compromised due to the potential failures of through-silicon-vias (TSVs) that are predominantly used as vertical interconnects in a 3D IC. To address these problems, we propose a machine-learning-inspired predictive design methodology for energy-efficient and reliable many-core architectures enabled by 3D integration. We demonstrate that a small-world network-based 3D NoC (3D SWNoC) performs significantly better than its 3D MESH-based counterparts. On average, the 3D SWNoC shows 35% energy-delay-product (EDP) improvement over 3D MESH for the PARSEC and SPLASH2 benchmarks considered in this work. To improve the reliability of 3D NoC, we propose a computationally efficient spare-vertical link (sVL) allocation algorithm based on a state-space search formulation. Our results show that the proposed sVL allocation algorithm can significantly improve the reliability as well as the lifetime of 3D SWNoC.