Large language model (LLM) inference demands vast memory capacity and hierarchical memory structures, but conventional architectures suffer from excessive energy and latency costs due to frequent data movement across deep memory tiers. To address this, we propose a wafer-scale SRAM-based Computing-in-Memory (CIM) architecture that performs all LLM operations in situ within the first-level SRAM, eliminating off-chip data migration and achieving unprecedented energy efficiency. However, wafer-scale SRAM CIM presents multiple challenges due to the limited first-level memory capacity, which requires efficient compute-memory resource allocation.

Graph processing suffers from severe locality challenges due to considerable inefficient irregular memory accesses, which mainly originate from random accesses to neighbors of active vertices (a.k.a frontiers). Graph reordering, which assigns continuous IDs to vertices that are more likely to be accessed consecutively, can improve access locality effectively and has demonstrated significant speedups across various architectures and systems. Existing graph reordering methods primarily explore the overlapping intensity of in-neighbor vertices for the data access locality characterization. However, many graph algorithms often activate a fraction of the vertices across the graphs, which vary substantially over different inputs and processing iterations. Many of these vertices are neither connected nor have any shared neighbors, but they are actually processed at the same time and exhibit potential data access locality, which is generally overlooked in prior graph reordering methods. We notice that the data locality between concurrently activated vertices are usually attributed to the overlapped$k$-order in-neighbors. As the number of$k$-order in-neighbors grows explosively, it is unacceptably time-consuming to analyze the overlapping of$k$-order in-neighbors for graph reordering directly. In this case, we propose to replace the overlapping calculation of$k$-order in-neighbors with frontier distribution analysis of a few BFS samplings. Specifically, we profile the frontiers distributed across iterations of different BFS samplings first and build a feature vector based on the activated iteration order of each vertex in the BFS samplings. On top of the feature vectors, we propose FrontOrder, which has a customized distance metric to characterize the locality between different vertices and leverages$K$-means to cluster vertices with high locality to guide graph reordering. In addition, FrontOrder also takes the load balance into consideration by predicting the runtime computing intensity with the learned clusters of vertices. According to our experiments, FrontOrder delivers an average performance speedup of${2.33\times}$and${1.57\times}$on Ligra and GPOP, respectively, and consistently outperforms the state-of-the-art graph reordering methods on a set of representative graph algorithms and datasets with moderate preprocessing overhead.

Out-of-core graph processing systems are severely bottlenecked by I/O to the external storage because of the low compute-to-I/O ratio and the substantial amount of irregular data accesses. In order to alleviate the I/O bottleneck, prior works either focus on improving the bandwidth utilization by converting random I/O requests into sequential ones, or improving the data utilization by fetching only the required data to avoid the I/O redundancy. However, the former usually loads massive unused data, while the latter can induce frequent finegrained I/O requests, wasting the parallelism of the I/O channels and leading to under-utilization of the limited I/O bandwidth. Different from prior works, we systematically explore the use of computational storage devices (CSDs), which offer in-storage computing facilities with higher I/O bandwidth, to improve both the bandwidth utilization and data utilization for higher I/O efficiency. Specifically, we first introduce a graph-semanticaware data organization to enable the loading of only active graph partitions at the granularity of a physical page, reducing redundant I/O and enhancing data utilization. Additionally, we propose to coalesce parallel I/O requests of graph partitions distributed across different flash dies to maximize the parallelism of internal I/O channels, thereby fully utilizing the internal I/O bandwidth of CSDs. In addition, we capture the dynamic status of graph processing tasks across the iterations and partitions at runtime to dynamically offload I/O-intensive workloads into the instorage processors with restricted computing resources but higher I/O bandwidth to further improve the I/O efficiency. With the above techniques, we implement an out-of-core graph processing system prototype, namely TaijiGraph, on an open-channel CSD. According to our experiments on a set of representative graph datasets and algorithms, TaijiGraph achieves average speedups of$2.43 \times, 3.81 \times, 2.21 \times$and$7.89 \times$, respectively, when compared to state-of-the-art out-of-core graph processing systems including GridGraph, LUMOS, Blaze, and GraphSSD.

Due to hardware customization capabilities, FPGA-based graph processing accelerators achieve significantly higher energy efficiency than many general-purpose computing engines. However, designing these accelerators remains a substantial challenge for high-level users. To overcome the programming barrier, FPGA-based accelerator design frameworks on top of generic graph processing programming models have been developed to automate accelerator generation through pre-built templates. However, they often tightly couple graph processing algorithms, programming models and processing paradigms, and accelerator architectures, which severely limits the expression scope of the algorithms and may also restrict the performance when the generated accelerators fail to suit dynamic processing patterns of the graph processing algorithms.

Graph reordering is an effective technique for improving the access locality of graph processing. However, existing methods often overlook the data access locality among concurrently activated vertices (a.k.a. frontiers). These vertices, while lacking direct connections or shared neighbors, can exhibit significant locality attributed to their overlapped k-order in-neighbors. However, calculating such overlaps directly is computationally prohibitive. We propose to estimate the overlapped k-order in-neighbors through frontier distribution analysis based on a few BFS samples. Our proposed graph reordering method, FrontOrder, constructs feature vectors from the frontier distribution of BFS samples, and employs K-means clustering with a custom distance metric to group vertices with high locality. Additionally, the learned clusters can predict runtime computing intensity, enabling load balancing through vertex reordering. FrontOrder achieves average speedups of 2.65× on Ligra and 1.73× on GPOP, outperforming state-of-the-art methods.

Approximate nearest neighbor search (ANNS) is essential for applications like recommendation systems and retrieval-augmented generation (RAG) but is highly I/O-intensive and memory-demanding. CPUs face I/O bottlenecks, while GPUs are constrained by limited memory. DRAM-based Processing-in-Memory (DRAM-PIM) offers a promising alternative by providing high bandwidth, large memory capacity, and near-data computation. This work introduces DRIM-ANN, the first optimized ANNS engine leveraging UPMEM’s DRAM-PIM. While UPMEM scales memory bandwidth and capacity, it suffers from low computing power because of the limited processor embedded in each DRAM bank. To address this, we systematically optimize ANNS approximation configurations and replace expensive squaring operations with lookup tables to align the computing requirements with UPMEM’s architecture. Additionally, we propose load-balancing and I/O optimization strategies to maximize parallel processing efficiency. Experimental results show that DRIM-ANN achieves a 2.46× speedup over a 32-thread CPU and up to 2.67× over a GPU when deployed on computationally enhanced PIM platforms.

Hardware-aware Neural Architecture Search (HW-NAS) has garnered significant research interest due to its ability to automate the design of neural networks for various hardware platforms. Prevalent HW-NAS frameworks often use fast predictors to estimate network performance, bypassing the time-consuming actual profiling step. However, the resource-intensive nature of building these predictors and their accuracy limitations hinder their practical use in diverse deployment scenarios. In response, we emphasize the indispensable role of actual profiling in HW-NAS and explore efficiency optimization possibilities within the HW-NAS framework. We provide a systematic analysis of profiling overhead in HW-NAS and identify many redundant and unnecessary operations during the search phase. We then optimize the workflow and present Insitu NAS, which leverages similarity features and exploration history to eliminate redundancy and improve runtime efficiency. In-situ NAS also offers simplified interfaces to ease the user’s effort in managing the complex device-dependent profiling flow, enabling plug-and-search functionality across diverse hardware platforms. Experimental results show that In-situ NAS achieves an average 10x speedup across different hardware platforms while reducing the search overhead by 8x compared to predictor-based approaches in various deployment scenarios. Additionally, In-situ NAS consistently discovers networks with better accuracy (about 1.5%) across diverse hardware platforms compared to predictor-based NAS.

Resistive random access memory (RRAM) emerges as a promising technology for developing energy-efficient deep neural network (DNN) accelerators, owing to its analog computing paradigm for matrix-vector multiplication. However, the inherent nonideal device features of RRAM cells, such as device variation, read disturbances, and limited on/off ratio, present challenges for model deployment. Therefore, to ensure accurate storage and computing precision for RRAM-based accelerators, a widely used practice is encoding a DNN weight by multiple cells, resulting in significant memory overhead and underutilization. This challenge is further exacerbated by the rapid increases in model size witnessed in recent years. While the one-to-one weight-cell mapping strategy can improve memory utilization, it inevitably introduces deviations in the mapped DNN weight from the desired value due to RRAM variation issues, leading to model accuracy degradation. In response to this challenge, we abstract the model optimization on RRAM chips as a non-ideal PIM device optimization problem, aimed at optimizing model accuracy without the requirement of precise weight programming. We systematically analyze the model optimization behavior on multilevel RRAM devices by investigating the accuracy recovery process of various fine-tuning strategies in recovering model performance under the non-ideal PIM device setting. Based on the analysis, we propose a non-ideal PIM device finetune scheme to recover the model performance for multilevel RRAM under the non-ideal PIM device setting. Our proposed scheme leverages knowledge distillation and exploits input/output information of the model on RRAM to guide the fine-tuning process, finally restoring its accuracy. Experimental results demonstrate the efficacy of our non-ideal PIM device fine-tuning scheme, achieving nearly complete recovery of model performance. Our approach yields over a 3% improvement in model accuracy compared to variation-aware training approaches.

In recent years, various computing-in-memory (CIM) processors have been presented, showing superior performance over traditional architectures. To unleash the potential of various CIM architectures, such as device precision, crossbar size, and crossbar number, it is necessary to develop compilation tools that are fully aware of the CIM architectural details and implementation diversity. However, due to the lack of architectural support in current popular open-source compiling stacks such as TVM, existing CIM designs either manually deploy networks or build their own compilers, which is time-consuming and labor-intensive. Although some works expose the specific CIM device programming interfaces to compilers, they are often bound to a fixed CIM architecture, lacking the flexibility to support the CIM architectures with different computing granularity. On the other hand, existing compilation works usually consider the scheduling of limited operation types (such as crossbar-bound matrix-vector multiplication). Unlike conventional processors, CIM accelerators are featured by their diverse architecture, circuit, and device, which cannot be simply abstracted by a single level if we seek to fully explore the advantages brought by CIM.

Graph is widely utilized as a key data structure in many applications such as social network and recommendation systems. However, many real-world graphs are constructed with large-scale correlated data, which typically involves massive random memory accesses and distance calculation resulting in considerable processing time and energy consumption on CPUs and GPUs. In this work, we present GCiM, a specialized processing-in-memory architecture for efficient graph construction and updating. By directly deploying the computing units on the logic layer of the 3D stacked memory, GCiM benefits from memory-level parallelism and further improves the memory access efficiency with both optimized processing ordering and data layout. In addition, we notice that the computing engines for graph updating suffer from dramatic utilization imbalance and integrate a power gating module to cut down the power supply of the idle computing engines at runtime and further enhance the energy efficiency. According to our experiments, GCiM shows 634.64X and 56.27X speedup while consuming 1194.14X and 505.07X less energy compared to CPU and GPU respectively.

Dynamic graphs can be utilized to model many real-world applications like social media analysis in which the connections and entities evolve continuously. Hence, the processing of dynamic graphs is gaining increasing popularity. However, prior dynamic graph processing systems mainly focus on the optimization of graph analytics but overlook graph updating which manages the evolving graph structure and presents a unified view to graph analytics. Since graph updating operates on evolving graphs and involves a large number of irregular memory accesses, it poses a substantial influence on the performance of dynamic graph processing systems. In this work, we observe that graph updating is mainly bottlenecked by a frequent indirect memory access pattern *(*(BAi+offset)). The pattern is inherent to the typical graph updating from the incoming edge stream to the base data store organized with either an adjacent list or a compressed sparse row. With this observation, we propose a novel Prefetcher for Dynamic Graph updating abbreviated as PDG. PDG is a lightweight pipelined instruction-based prefetcher specialized for graph updating and it is also compatible with the irregular memory access pattern BAi widely used in graph analytics. In addition, it leverages a monitor of the instruction queue to decide the appropriate timing of prefetching to make the best use of the cache. According to our experiments, PDG achieves 1.60×, 1.26× and 1.30× performance speedup compared to three representative prefetchers respectively with negligible hardware overhead in graph updating.

In this work, we propose Layer-Puzzle, a multi-task allocation and scheduling framework for multi-core NPUs. Based on the proposed latency-prediction model and dynamic parallelization scheme, Layer-Puzzle can generate near-optimal results for each layer under given hardware resources and traffic congestion levels. As an online scheduler, Layer-Puzzle performs a QoS-aware and dynamic scheduling method that picks the superior version from the previously compiled results and co-runs the selected tasks to improve system performance. Our experiments on MLPerf show that Layer-Puzzle can achieve up to 1.61X, 1.53X, and 1.95X improvement in ANTT, STP, and PE utilization, respectively.

Compared to conventional neural networks, training a supernet for Neural Architecture Search (NAS) is very time consuming. Although current works have demonstrated that parallel computing can significantly speed up the training process, almost all of their parallelism still follow the conventional data- and model-based paradigms, which actually face performance issues in both computation and inter-node communication of the supernet training. To further improve the performance of current methods, we discover the unique path-parallelism that exists in supernet training, and proposed a novel training approach designed specifically for supernet. In detail, we focus on analyzing path correlations between subnets in a supernet and exploiting effective path-merging methods to reduce redundant computations and communications raised by concurrent subnets. Moreover, we also try to combine the proposed path parallelism with traditional intra-subnet parallelism to perform multi-level parallelization to further optimize the parallel performance. We present the detailed design and implementation of our method, and our experimental results show that our proposed approach can achieve up to 3.2x end-to-end speedup over conventional parallel training solutions, and 1.46x–5.78x speedup compared to the state-of-art supernet training frameworks.

Recently, video processing tasks, such as video enhancement and analysis, have received increasing attention from both academics and industries. However, the current video processing procedure on edge decouples the decoding phase and the subsequent video processing tasks, missing the opportunity to accelerate the procedure by orchestrating video decoding and enhancement stages. Thus, we propose an intelligent video processing workflow and architecture(IVP) for cloud-edge video streaming. For edge devices that receive compressed videos, IVP can perform direct DNN-based video enhancement, e.g., super-resolution and frame-interpolation. By leveraging the metadata motion vectors and residuals extracted from the encoded video, our architecture will significantly eliminate unnecessary frame pixels being processed by the DNNs and improve execution efficiency. The proposed IVP and workflow are proved to reduce up to 90% of the processing latency while producing accurate and high-quality videos. Furthermore, we observe a significant portion of similar optical flow in time domain of continuous videos, which can be used to reduce the computation overhead. Thus, to utilize such temporal similarity of optical flow, the proposed IVP is upgraded to be capable of reusing previous computation results, which further improves energy efficiency of the whole system.

The RRAM-based neuromorphic computing system (NCS) has amassed explosive interests for its superior data processing capability and energy efficiency than traditional architectures, and thus being widely used in many data-centric applications. The reliability and security issues of the NCS, therefore, become an essential problem. In this article, we systematically investigated the adversarial threats to the RRAM-based NCS and observed that the RRAM hardware feature can be leveraged to strengthen the attack effect, which has not been granted sufficient attention by previous algorithmic attack methods. Thus, we proposed two types of hardware-aware attack methods with respect to different attack scenarios and objectives. The first is an adversarial attack, VADER, which perturbs the input samples to mislead the prediction of neural networks. The second is fault injection attack, EFI, which perturbs the network parameter space such that a specified sample will be classified to a target label, while maintaining the prediction accuracy on other samples. Both attack methods leverage the RRAM properties to improve the performance compared with the conventional attack methods. Experimental results show that our hardware-aware attack methods can achieve nearly 100% attack success rate with extremely low operational cost, while maintaining the attack stealthiness.

Network sparsity or pruning is a pivotal technology for edge intelligence. Resistive random access memory (RRAM)-based accelerators, featuring dense storage and processing in memory capability, have demonstrated the superior computing performance and energy efficiency over the traditional CMOS-based accelerators for neural network applications. Unfortunately, RRAM-based accelerators suffer the performance or energy degradation when deploying pruned models, impairing their competition in the edge intelligence scenarios. We observed the essential reason is the pruning technology and the mapping strategy in prior RRAM-based accelerator and are optimized individually. As a result, the random zeros in the pruned deep neural network are irregularly distributed in the crossbars, rendering the degradation of computation parallelism of the crossbar without crossbar demand reduction. In this work, we propose a coordinated model pruning and mapping framework to jointly optimize of model accuracy and efficiency of RRAM-based accelerators. As for the mapping, we first decouple weight matrices in the bit-wise way and map the bit matrices to different crossbars, where the signed weights are represented with the two’s complement so as that save half desired crossbars. As for the pruning, we prune weight bits at the crossbar granularity so that free the crossbars holding the pruned bits. Furthermore, we employ an reinforcement learning (RL) approach to automatically select the optimal crossbar-aware bit-pruning strategy for any given neural network without laborious human efforts. We conducted the experiments on a set of representative neural networks and compared our framework with the state-of-the-art (SOTA) bit-sparsity works. The results show that automatic structured bit-pruning saves up to 89.64% energy reduction and 84.12% area overhead compared to existing PRIME-like architecture. Besides, our framework outperforms the SOTA bit-sparsity design by$1.5\times $in terms of the energy reduction on the RRAM-based accelerator.

Bit-serial architectures (BSAs) are becoming increasingly popular in low-power neural network processor (NNP) designs for edge scenarios. However, the performance and energy efficiency of state-of-the-art BSA NNPs heavily depends on both the proportion and distribution of ineffectual weight bits in neural networks (NNs). To boost the performance of typical BSA accelerators, we present Bit-Pruner, a software approach to learn BSA-favored NNs without resorting to hardware modifications. Bit-Pruner not only progressively prunes but also restructures the nonzero bits in weights so that the number of nonzero bits in the model can be reduced and the corresponding computing can be load-balanced to suit the target BSA accelerators. On top of Bit-Pruner, we further propose a Pareto frontier optimization algorithm to adjust the bit-pruning rate across network layers and fulfill diverse NN processing requirements in terms of performance and accuracy for various edge scenarios. However, an aggressive Bit-Pruner can lead to nontrivial accuracy loss, especially for lightweight NNs and complex tasks. To this end, the alternating direction method of multipliers (ADMMs) is adapted to the retraining phase in Bit-Pruner to smooth the abrupt disturbance due to bit-pruning and enhance the resulting model accuracy. According to the experiments, Bit-Pruner increases the bit-sparsity up to 94.4% with negligible accuracy degradation and achieves an optimized tradeoff between NN accuracy and energy efficiency even under very-aggressive performance constraints. When pruned models are deployed onto typical BSA accelerators, the average performance is$2.1\times $and$1.6\times $higher than the baseline networks without pruning and those with classical weight pruning, respectively.

The reduction of neural parameters and operations for the applications on embedded and IoT platforms in current deep neural network (DNN) architectures has received increasing attention. Relatively, the intermediate feature maps of such lightweight neural networks begin to grow and usually outsize the on-chip memory as the new bottleneck, which introduces considerable power-consuming off-chip memory accesses. To reduce the feature-induced memory accesses, operator fusion has been proposed to parallelize the execution of multiple convolutional layers and shown significant reduction of off-chip memory accesses. However, how to fuse the neural operators is still a challenging issue that heavily depends on both the neural network (NN) topology and the specific DNN accelerator configuration. In this work, we observed prior operator fusion approaches fail to guarantee memory-level optimality as they search in the constrained operator fusion design space. Considering the complexity of the NN topologies and the constrained resources of the DNN accelerators, we develop a novel operator fusion framework, Optimus. Optimus includes an accurate memory cost model dedicated to the scheduler to evaluate the potential operator-fusion schemes and a directed acyclic graph-based operator fusion algorithm for both off-line and on-line workload deployment scenarios, which altogether generates high-efficiency operator-fusion solutions for arbitrary network models running on DNN accelerators. The experimental results show that Optimus reduces 17–75% off-chip memory accesses and obtains 1.86×–3.66× energy efficiency on state-of-the-art DNN workloads when compared to the baselines and brings significant power-efficiency boost to the DNN accelerators of different architectures and dataflows.

The efficiency of deep neural network (DNN) solutions on real hardware devices are mainly decided by the DNN architecture and the compiler-level scheduling strategy on the hardware. When we try to fully exploit the underlying hardware and obtain the optimal tradeoff between DNN accuracy and runtime performance, we discovered that the two optimization goals of DNN architecture and scheduling policy are intimately related to each other. However, current hardware-aware Neural Architecture Search (NAS) methods primarily focus on the DNN architecture search process, ignoring the effects of various compiler-level scheduling strategies (e.g., graph-level optimization, loop transformations, parallelization, etc.) on network candidates being evaluated in the search process. As a result, they may overlook the true-optimal DNN implementations on hardware, which can only be discovered by trying-out different combinations of scheduling strategies and DNN architectures. This work proposes a NAS framework (CHaNAS) that searches for not only the network architecture but also the dedicated compiler-level scheduling policy, as the optimal co-design solution on the target hardware. We propose to use a block-based pre-scheduling methodology to reduce the co-design search space and enable the automatic generation of the optimal co-design, including the network architecture and the tensor programs that practice the scheduling policy. Further, we introduce a new search objective function based on the generalization gap to prevent the selection of architectures that are prone to overfitting. We evaluate CHaNAS on Imagenet on different hardware back-ends against the state-of-the-art hardware-aware search method based on the MobileNet-v3 search space. Experimental results show that the co-design solutions obtained by ChaNAS show up to 1.6×, 1.9×, and 1.7×, 24 performance boost on NVIDIA P100 GPU, Intel Xeon 8163 CPU, and Samsung Note 10 Mobile, respectively, over the baselines of the same-level accuracy.

Winograd is generally utilized to optimize convolution performance and computational efficiency because of the reduced multiplication operations, but the reliability issues brought by winograd are usually overlooked. In this work, we observe the great potential of winograd convolution (WG-Conv) in improving neural network (NN) fault tolerance. Based on the observation, we evaluate WG-Conv fault tolerance comprehensively from different granularities ranging from models, layers, and operation types for the first time. Then, we explore the use of inherent fault tolerance of WG-Conv for cost-effective NN protection against soft errors. Specifically, we mainly investigate how WG-Conv can be effectively incorporated with classical fault-tolerant design approaches including triple modular redundancy (TMR), fault-aware retraining, and constrained activation functions. According to our experiments, WG-Conv can reduce the fault-tolerant design overhead by 55.77% on average without any accuracy loss compared to standard convolution (ST-Conv), and further reduce the computing overhead by 17.24% when the inherent fault tolerance of WG-Conv is considered. When it is applied on fault-tolerant NNs enhanced with fault-aware retraining and constrained activation functions, the resulting model accuracy generally shows significant improvement in the presence of various faults.

Vision transformers (ViTs) that leverage self-attention mechanism have shown superior performance on many classical vision tasks compared to convolutional neural networks (CNNs) and gain increasing popularity recently. Existing ViTs’ works mainly optimize performance and accuracy, but ViTs’ reliability issues induced by soft errors in large-scale VLSI designs have generally been overlooked. In this work, we mainly study the reliability of ViTs and investigate the vulnerability from different architecture granularities ranging from models, layers, modules, and patches for the first time. The investigation reveals that ViTs with the self-attention mechanism are generally more resilient on linear computing including general matrix–matrix multiplication (GEMM) and full connection (FC) and show a relatively even vulnerability distribution across the patches. ViTs involve more fragile non-linear computing such as softmax and GELU compared to typical CNNs. With the above observations, we propose a lightweight block-wise algorithm-based fault-tolerance (LB-ABFT) approach to protect the linear computing implemented with distinct sizes of GEMM and apply a range-based protection scheme to mitigate soft errors in non-linear computing. According to our experiments, the proposed fault-tolerant approaches enhance ViTs’ accuracy significantly with minor computing overhead in the presence of various soft errors.

Real-time ego-motion tracking and 3D structural estimation are the fundamental tasks for the ubiquitous cyper-physical systems, and they can be conducted via the state-of-the-art Edge-Based Visual Odometry (EBVO) algorithm. However, the intrinsic data-intensive process of EBVO emplaces a memory-wall hurdle in practical deployment on conventional von-Neumann-style computing systems. In this work, we attempt to leverage SRAM based processing-in-memory (PIM) technique to alleviate such memory-wall bottleneck, so as to optimize the EBVO systematically from the perspectives of the algorithm layer and physical layer. In the algorithm layer, we first investigate the data reuse patterns of the essential computing kernels required for the feature detection and pose estimation steps in EBVO, and propose PIM friendly data layout and computing scheme for each kernel accordingly. We distill the basic logical and arithmetical operations required in the algorithm layer, and in the physical layer, we propose a novel bit-parallel and reconfigurable SRAM-PIM architecture to realize the operations with high computing precision and throughput. Our experimental result shows that the proposed multi-layer optimization allows for high tracking accuracy of EBVO, and it can improve 11x processing speed and reduce 20x energy consumption compared to the CPU implementation.

Winograd convolution is originally proposed to reduce the computing overhead by converting multiplication in neural network (NN) with addition via linear transformation. Other than the computing efficiency, we observe its great potential in improving NN fault tolerance and evaluate its fault tolerance comprehensively for the first time. Then, we explore the use of fault tolerance of winograd convolution for either fault-tolerant or energy-efficient NN processing. According to our experiments, winograd convolution can be utilized to reduce fault-tolerant design overhead by 27.49% or energy consumption by 7.19% without any accuracy loss compared to that without being aware of the fault tolerance.

The growing complexity and diversity of deep neural network (DNN) applications have inspired intensive research on specialized DNN accelerators and also the design automation frameworks. Previous specialized NN acceleratos roughly fall into two categories of implementation, either the no-pipelined architecture that relies on a generic processing unit (PU) to sequentially execute the DNN layers in a layer-wise way, or the fully-pipelined architecture that dedicates interconnected customized PUs to the corresponding DNN layers in the model. Thus, such designs often suffer from either the resource under-utilization issue faced by no-pipelined accelerators or the resource scalability problem brought by the over-deep pipeline designs. In this work, we propose a novel class of design solution for DNN acceleration, segment-grained pipeline architecture (SPA). In the SPA accelerator, the targeted workload of DNN models will be divided into many segments and each segment will be sequentially executed on the shared interconnected PUs in a pipeline manner, so that they will benefit from both the efficiency of pipelined execution and also the flexibility of sharing PUs across different model layers. Particularly, we found that the efficiency of the implemented SPA accelerator significantly depends on the segmentation strategies of the models and the hardware resources assignment policy for PUs. Therefore, we introduce an automated design framework, AutoSeg, that includes a parameterized SPA accelerator template and a co-design engine that will generate the efficient model segmentation solution and hardware pipeline design parameters for the acceleration workload. Experimental results show that the SPA solutions generated by the AutoSeg framework achieve $1.2\times to 6.3\times$ speedup when compared to ASIC-based general DNN processors, and the FPGA designs implemented by AutoSeg also achieve as high as $3.4\times$ DSP efficiency and $3.6\times$ throughput improvement.

In DNN processors, main memory consumes much more energy than arithmetic operations. Therefore, many memory-oriented network scheduling (MONS) techniques are introduced to exploit on-chip data reuse opportunities and reduce accesses to memory. However, to derive the theoretical lower bound of memory overhead for DNNs is still a significant challenge, which also sheds light on how to reach memory-level optimality by means of network scheduling. Prior work on MONS mainly focused on disparate optimization techniques or missed some of the data reusing opportunities in diverse network models, thus their results are likely to deviate from the true memory-optimality that can be achieved in processors. This paper introduces Olympus, which comprehensively considers the entire memory-level DNN scheduling space, formally analyzes the true memory-optimality and also how to reach the memory-optimal schedules for an arbitrary DNN running on a DNN processor. The key idea behind Olympus is to derive a true memory lower-bound regarding both the intra-layer and inter-layer reuse opportunities, which has not been simultaneously explored by prior works. Evaluation on SOTA DNN processors of different architectures shows that Olympus can guarantee the minimum off-chip memory access, and it reduces 12.3-85.6% DRAM access and saves 7.4-70.3% energy on the latest network models.

Dynamic resources management in reconfigurable processors often manifests as a hard online decision-making task, which should yield premier solutions that must meet Quality-of-Service (QoS) requirements while maximizing the system’s efficiency. Most prior works rely on a hard-to-train predictor to model the complicated relationships between processor configurations and performance. To decide the proper resource allocation, the predictor needs to tentatively evaluate a group of possible configurations, and then decide the best configuration for the workload. This tedious process has an expensive runtime overhead for resource configuration in processors. Besides, prior works focus on improving the prediction accuracy, however, higher performance prediction cannot guarantee a good system outcome. Inspired by recent advances in adversarial learning, we present a generative adversarial network (GAN)-based framework, Amphis, which can directly generate the on-demand processor configuration for any scheduled-in application. By evaluating Amphis on a reconfigurable processor with 18 different workloads, our results demonstrate that the GAN-based method provides tremendous overhead reduction (up to 90%) compared to the SOTA prediction-based method WNNM while providing higher resource utilization.

Adversarial attacks have been viewed as the primary threat to the security of neural networks. Hence, extensive adversarial defense techniques have been proposed to protect the neural networks from adversarial attacks, allowing for the application of neural networks to the security-sensitive tasks. Recently, the emerging devices, e.g., Resistive RAM (RRAM), attracted extensive attention for establishing the hardware platform for neural networks to tackle the inadequate computing capability of the traditional computing platform. Though the emerging devices exhibit the instinct instability issues due to the advanced manufacture technology, including hardware variations and defects, the error-resilience capability of neural networks enables the wide deployment of neural networks on the emerging devices. In this work, we find that the natural instability in emerging devices impairs the security of neural networks. Specifically, we design an enhanced adversarial attack, Variation-oriented ADvERsarial (VADER) attack which leverages the inherent hardware variations in RRAM chips to penetrate the protection of adversarial defenses and mislead the prediction of neural networks. We evaluated the effectiveness of VADER across various protected neural network models and the result shows that VADER achieves higher success attack rate over other adversarial attacks.

This work proposes an intelligent video processing architecture for bandwidth-efficient edge-cloud video streaming. On receiving the bandwidth-saving low-quality video streaming in compressed format, the proposed architecture can perform direct DNN-based video enhancement, e.g., super-resolution and motion-compensated frame interpolation (MCFI), on streams. By utilizing the metadata motion vectors and residuals extracted from the encoded video, our workflow will significantly eliminate the unnecessary pixels being processed by the video-enhancing DNNs, and greatly promote the execution efficiency. The evaluation results on popular datasets show that our architecture can reduce the edge-side processing latency of video-enhancing DNNs by 90% compared to the traditional flow while producing accurate and high-quality videos on edge.

Network sparsity or pruning is an extensively studied method to optimize the computation efficiency of deep neural networks (DNNs) for CMOS-based accelerators, such as FPGAs and GPUs. Though the RRAM-based accelerator has demonstrated superior performance and energy efficiency for DNN tasks, deploying the sparse neural networks desires dedicated consideration to save resource consumption without introducing the expensive index overhead and sophisticated control. To exploit the potential of sparse neural network design on the RRAM-based accelerator, we propose an automatic structured bit-pruning design, ASBP, to harmonize the optimization objective of DNN sparsity with efficient RRAM deployment. Specifically, ASBP prunes the bits of weight which are split into different crossbars and thus, free the zero-value crossbar when mapping the neural network into RRAM-based accelerators without extra hardware modification. Meanwhile, ASBP employs the reinforcement learning (RL) approach to automatically select the best crossbar-aware bit-sparsity strategy for any given neural network without laborious human efforts. According to our experiments on a set of representative neural networks, ASBP saves up to 79.01% energy consumption and 54.79% area overhead compared to the baseline that deploys the original DNN on the RRAM-based accelerator. Besides, ASBP outperforms the state-of-the-art bit-sparsity design by 1.4x in terms of the energy reduction on the RRAM-based accelerator.

Chiplet based multi-die integration has been thought as a key enabler of the agile chip development flow. For 2.5D based multi-die system, Network on Interposer plays an essential role in the performance and the development cost of the chips. This work proposed a reusable NoI design for agile AI chip customization. The proposed NoI design can self-adapt to the inter-die communication patterns of various neural network applications, so the produced interposers can be reused across different AI chip specifications. Experimental results show the proposed NoI design brings 42.7%$\sim$79.5% of total data communication latency reduction in different scenarios, and it also decreased the area overhead by 26.4%.

As neural networks (NNs) are being widely deployed in many cloud-oriented systems for safety-critical tasks, the privacy and security of NNs become significant concerns to users in the cloud platform that shares the computation infrastructure such as memory resource. In this work, we observed that the memory timing channel in the shared memory of cloud multi-core architecture poses the risk of network model information leakage. Based on the observation, we propose a learning-based method to steal the model architecture of the NNs by exploiting the memory timing channel without any high-level privilege or physical access. We first trained an end-to-end measurement network offline to learn the relation between memory timing information and NNs model architecture. Then, we performed an online attack and reconstructed the target model using the prediction from the measurement network. We evaluated the proposed attack method on a multi-core architecture simulator. The experimental results show that our learning-based attack method can reconstruct the target model with high accuracy and improve the adversarial attack success rate by 42.4%.

Multi-task learning systems are commonly adopted in many real-world AI applications such as intelligent robots and self-driving vehicles. Instead of improving single-network performance, this work proposes a specialized Multi-Task Deep Learning Accelerator architecture, MT-DLA, to improve the performance of concurrent networks by exploiting the shared feature and parameters across these models. It is shown in our evaluation with realistic multi-task workloads, MT-DLA dramatically eliminates the memory and computation overhead caused by the shared parameters, activations and computation result. In the experiments with real-world multi-task learning workloads, MT-DLA brings about 1.4x-7.0x energy efficiency boost when compared to the baseline neural network accelerator without multi-task support.

Ego-motion estimation with 3D perception using visual odometry (VO) is known to be robust and economical among the existing odometry techniques. However, existing VO solutions are typically both computation intensive and memory intensive, which dramatically inhibits their deployment in IoT platforms such as robotic vehicles and handheld devices mostly equipped with resource-constrained MCU-level processors. To enable real-time and high-quality VO on these scenarios with thrifty resource budgets, we investigate state-of-the-art edge-based VO (EBVO) and propose an optimization framework called PicoVO that can greatly reduce the amount of computation as well as the memory footprint from the perspectives of both algorithm and implementation. First of all, we revisit the key processing stages of EBVO and propose an EBVO-oriented lightweight edge detector in the pre-processing stage, a sparse-to-dense processing scheme in the tracking stage, and a lightweight key-frame management in the post-processing stage. In addition to the algorithmic optimization, we further develop a dedicated quantization scheme particularly for the 3D feature calculation and Levenberg-Marquardt (LM) solver that are critical to the memory footprint and computation requirements of PicoVO. Evaluation on realistic RGB-D benchmark datasets is conducted on NUCLEO-F767ZI equipped with a 216MHz Cortex-M7 MCU and 512KB RAM. It reveals that PicoVO achieves 33fps@320x240 with high tracking precision comparable to state-of-the-art VOs on PC.

Neural network search (NAS) projects a promising direction to automate the design process of efficient and powerful neural network architectures. Nevertheless, the NAS techniques have to dynamically generate a large number of candidate neural networks, and iteratively train and evaluate these on-line generated network architectures, thus they are extremely time-consuming even when deployed on large GPU clusters, which dramatically hinders the adoption of NAS. Though recently there are many specialized architectures proposed to accelerate the training or inference of neural networks, we observe that existing neural network accelerators are typically targeted at static neural network architectures, and they are not suitable to accelerate the evaluation of the dynamical neural network candidates evolving during the NAS process, which cannot be deployed onto current accelerators via the off-line compilation.To enable rapid and energy-efficient NAS in compact single-chip solutions, we propose NASA, a specialized architecture for one-shot based NAS acceleration. It is able to generate, schedule, and evaluate the candidate neural network architectures for the target machine learning workload with high speed, significantly alleviating the processing bottleneck of one-shot NAS. Motivated by the observation that there are considerable computation sharing opportunities among the different neural network candidates generated in one-shot NAS, NASA is equipped with an on-chip network fusion unit to remove the redundant computation during the network mapping stage. In addition, the NASA accelerator can partition and re-schedule the candidate neural network architectures at fine-granularity to maximize the chance of data reuse and improve the utilization of the accelerator arrays integrated to accelerate network evaluation. According to our experiments on multiple one-shot NAS tasks, NASA achieves 33.52× performance speedup and 214.33× energy consumption reduction on average when compared to aCPU-GPU system.

Neural network layer fusion has been proposed to parallelize the inference of neural layers and thus significantly reduces the feature-induced memory accesses. However, how to fuse the neural layers is still a challenging issue that heavily depends on both the network architecture and the specific DNN processor configuration. This work formalizes the layer fusion problem for DNN processors, proves that prior fusion solutions cannot guarantee memory-level optimality, and presents a novel neural network fusion framework, Optimus. Optimus includes an accurate memory cost model to evaluate fusion schemes, and a Computing-Graph (CG) based layer fusion algorithm, which generates high-efficiency layer-fusion schemes for arbitrary network architectures on DNN processors. The proposed off-line and on-line graph-based fusion algorithms can reduce 10.1% - 72.2% off-chip memory traffic and obtain 1.71x - 3.94x energy efficiency over SOTA baselines on DNN workloads, and they bring significant power-efficiency boost to the DNN processors of different architectures and dataflows.

Automatically design an efficient DNN solution for a given deep learning task on the target hardware mainly decided by the neural network architecture and the schedule mapping strategy, where the two goals are closely coupled with each other to fully exploit the advantages of the underlying hardware. Prior hardware-aware Neural Architecture Search (NAS) methods mostly ignore the impacts of different scheduling policies (e.g., graph-level optimization, loop transformations, parallelization, etc.) on network candidates being evaluated in the search process. Thus, they may miss the true-optimal architecture that can only be discovered by trying-out different scheduling policies. This work proposes a NAS framework (CHaNAS) that searches for not only the network architecture but also the dedicated scheduling policy, as the optimal co-design solution on target hardware that fully exploits the advantages of the underlying hardware. We propose to use a block-based pre-scheduling methodology to reduce the co-design search space, and enable the automatic generation of the optimal co-design, including the network architecture and the tensor programs that practice the scheduling policy. We evaluate CHaNAS on Imagenet on different hardware back-ends against the state-of-the-art hardware-aware search method MobileNet-v3. Experimental results show that the co-design solutions obtained by ChaNAS show up to 1.6x, 1.9x, and 1.7x performance boost on NVIDIA P100 GPU, Intel Xeon 8163 CPU, and Samsung Note 10 Mobile, respectively, over the baselines of the same-level accuracy.

Prior works typically conducted the fault analysis of neural network accelerator computing arrays with simulation and focused on the prediction accuracy loss of the neural network models. There is still a lack of systematic fault analysis of the neural network acceleration system that considers both the accuracy degradation and system exceptions, such as system stall and running overtime. To that end, we implemented a representative neural network accelerator and corresponding fault injection modules on a Xilinx ARM-FPGA platform and evaluated the reliability of the system under different fault injection rates when a series of typical neural network models are deployed on the neural network acceleration system. The entire fault injection and reliability evaluation system is open-sourced on GitHub. With comprehensive experiments on the system, we identify the system exceptions based on the various abnormal behaviors of the FPGA-based neural network acceleration system and analyze the underlying reasons. Particularly, we find that the probability of the system exceptions dominates the reliability of the system. The faults also incur accuracy degradation of the neural network models, but the influence depends on the applications of the models and can vary greatly. In addition, we also evaluated the use of conventional triple modular redundancy (TMR) and demonstrated the challenge of TMR with both experiments and analytical models, which may shed light on the reliability design of the FPGA-based neural network acceleration system.

The increasing hardware failures caused by the shrinking semiconductor technologies pose substantial influence on the neural accelerators and improving the resilience of the neural network execution becomes a great design challenge especially to mission-critical applications such as self-driving and medical diagnose. The reliability analysis of the neural network execution is a key step to understand the influence of the hardware failures, and thus is highly demanded. Prior works typically conducted the fault analysis of neural network accelerators with simulation and concentrated on the prediction accuracy loss of the models. There is still a lack of systematic fault analysis of the neural network acceleration system that considers both the accuracy degradation and system exceptions such as system stall and early termination.In this work, we implemented a representative neural network accelerator and fault injection modules on a Xilinx ARM-FPGA platform and conducted fault analysis of the system using four typical neural network models. We had the system open-sourced on github. With comprehensive experiments, we identify the system exceptions based on the various abnormal behaviours of the FPGA-based neural network acceleration system and analyze the underlying reasons. Particularly, we find that the probability of the system exceptions dominates the reliability of the system and they are mainly caused by faults in the DMA, control unit and instruction memory of the accelerators. In addition, faults in these components also incur moderate accuracy degradation of the neural network models other than the system exceptions. Thus, these components are the most fragile part of the accelerators and need to be hardened for reliable neural network execution.

Convolutional neural networks (CNNs) have become the state-of-the-art technique in many classification tasks in IoT system. However, the low-power and area-constraint edge devices are unable to afford the expensive cost of CNNs. Resistive random access memory (RRAM) is attractive for establishing the CNN accelerator at the edge end due to the features of scalability, low-power and in-situ dot-product. However, mapping a random network architecture onto a general-purpose RRAM accelerator suffers a severe issue of resource underutilization. The neural network quantization offers an opportunity to rescue the degraded resource utilization. Selecting the bit-width for the vast parameters is impractically completed by human labor. This paper proposes an AutoML-based array-aware quantization and mapping framework that generates the fine-grained mixed-precision neural networks to optimize resource utilization in RRAM. In this framework, we design a two-stage learning and array-aware grouping strategy to quickly explore the huge searching space. The experimental results show that the proposed framework achieves 18.2%~36.1% improvement in resource utilization and 0.9%~3.3% increase in model accuracy over prior coarse-grained quantization methods.

Bit-serial architectures (BSAs) are becoming increasingly popular in low power neural network processor (NNP) design. However, the performance and efficiency of state-of-the-art BSA NNPs are heavily depending on the distribution of ineffectual weight-bits of the running neural network. To boost the efficiency of third-party BSA accelerators, this work presents Bit-Pruner, a software approach to learn BSA-favored neural networks without resorting to hardware modifications. The techniques proposed in this work not only progressively prune but also structure the non-zero bits in weights, so that the number of zero-bits in the model can be increased and also load-balanced to suit the architecture of the target BSA accelerators. According to our experiments on a set of representative neural networks, Bit-Pruner increases the bit-sparsity up to 94.4% with negligible accuracy degradation. When the bit-pruned models are deployed onto typical BSA accelerators, the average performance is 2.1X and 1.5X higher than the baselines running non-pruned and weight-pruned networks, respectively.

Carbon Nanotubu field-effect transistor (CNFET) that promises both higher clock speed and energy efficiency becomes an attractive alternative to the conventional power-hungry CMOS cache. We observe that the CNFET-based cache constructed with typical SRAM cells has distinct energy consumption when reading/writing 0 and 1 from/to it. For instance, the energy consumption of writing 1 to an SRAM cell is almost 10X higher than writing 0. With this observation, we propose an energy-efficient cache design called CNT-Cache to take advantage of this feature. It predicts the cache line access pattern based on the latest cache line access history. On top of the prediction, it decides the optimal cache line encoding to match the cache operation preferences at runtime. According to our experiments on a set of benchmark programs, the optimized CNFET-based D-Cache reduces the dynamic power consumption by 22.2% on average compared to the baseline CNFET cache.

DNN/Accelerator co-design has shown great potential in improving QoR and performance. Typical approaches separate the design flow into two-stage: (1) designing an application-specific DNN model with high accuracy; (2) building an accelerator considering the DNN specific characteristics. However, it may fails in promising the highest composite score which combines the goals of accuracy and other hardware-related constraints (e.g., latency, energy efficiency) when building a specific neural-network-based system. In this work, we present a single-stage automated framework, YOSO, aiming to generate the optimal solution of software-and-hardware that flexibly balances between the goal of accuracy, power, and QoS. Compared with the two-stage method on the baseline systolic array accelerator and Cifar10 dataset, we achieve 1.42x~2.29x energy or 1.79x~3.07x latency reduction at the same level of precision, for different user-specified energy and latency optimization constraints, respectively.

The design of neural network architecture for code approximation involves a large number of hyper-parameters to explore, it is a non-trivial task to find an neural-based approximate computing solution that meets the demand of application-specified accuracy and Quality of Service (QoS). Prior works do not address the problem of ‘optimal’ network architectures design in program approximation, which depends on the user-specified constraints, the complexity of dataset and the hardware configuration. In this paper, we apply Neural Architecture Search (NAS) for searching and selecting the neural approximate computing and provide an automatic framework that tries to generate the best-effort approximation result while satisfying the user-specified QoS/accuracy constraints. Compared with previous method, this work achieves more than 1.43x speedup and 1.74x energy reduction on average when applied to the AxBench benchmarks.

Processing-in-Memory (PIM) or Near-Data Processing has been recognized as the most potential solution to resolve the ever-aggravating memory wall especially as the thrive of memory-intensive scale-out workloads such as graph computing and data analytics. However, when the future computing system becomes more and more likely to adopt PIM architectures as a type of the storage and processing component, there is a lack of literature and research work on the general scheduling framework with the emerging heterogeneous system except for some ad-hoc task partitioning methods with specialized PIM designs. This work is the first to propose a formalized model to quantitatively describe the multi-task scheduling problem in PIM+CPU platform without loss of generality, and also an optimized task mapping-and-scheduling algorithm to boost the hardware utility for these novel heterogeneous systems. The proposed scheduling framework is fully aware of the data access bandwidth and processing capability distinction between the CPU and PIM devices, and also the implications of task mapping on the bandwidth contention, data communication intensity and hardware utility for the concurrent workloads. Experimental results show that, compared to the traditional scheduling algorithm for heterogeneous system, the proposed method is able to improve the system performance by over 10% and the energy efficiency by almost 10% for multi-core scale-out applications.

Regular 2D computing array is widely utilized for the processing of the major neural network operations in many deep learning accelerators (DLAs). Hardware failures on the array can lead to considerable computing errors and prediction accuracy loss. Prior works proposed to add homogeneous redundant PEs to each row or column of the regular computing array to mitigate faulty PEs, but they may fail to recover the computing array from faults when the number of faulty PEs in a row or column exceeds the number of redundant PEs in the corresponding row or column. The problem gets worse when the faults are not evenly distributed across the computing array. To address the problem, we propose a hybrid computing architecture (HCA) for fault-tolerant DLAs. Instead of adding homogeneous redundant PEs to the regular computing array of DLAs, it has a dot-production processing unit (DPPU) to recompute the operations that are mapped to the faulty PEs concurrently without performance penalty under moderate fault injection. Even under high fault injection, HCA can be degraded smoothly and remains functional. In addition, DPPU exploits the parallelism within each operation and processes the network operations sequentially, so it can tolerate faulty PEs in arbitrary locations and ensures steady performance under distinct fault distributions. According to our experiments, HCA shows significantly higher reliability and performance under various fault injection with comparable chip area penalty compared to the conventional redundancy approaches.

Generative neural network is a new category of neural networks and it has been widely utilized in many applications such as content generation, unsupervised learning, segmentation, and pose estimation. It typically involves massive computing-intensive deconvolution operations that cannot be fitted to conventional neural network processors directly. However, prior works mainly investigated specialized hardware architectures through intensive hardware modifications to the existing deep learning processors to accelerate deconvolution together with the convolution. In contrast, this article proposes a novel deconvolution implementation with a software approach and enables fast and efficient deconvolution execution on the existing deep learning processors. Our proposed method reorganizes the computation of deconvolution and allows the deep learning processors to treat it as the standard convolution by splitting the original deconvolution filters into multiple small filters. Compared to prior acceleration schemes, the implemented acceleration scheme achieves 2.4× -4.3× performance speedup and reduces the energy consumption by 27.7 -54.5 percent on a set of realistic benchmarks. In addition, we have also applied the deconvolution computing approach to the off-the-shelf commodity deep learning processors. The performance of deconvolution also exhibits significant performance speedup over prior deconvolution implementations.

Continuous-valued deep convolutional networks (DNNs) can be converted into accurate rate-coding based spike neural networks (SNNs). However, the substantial computational and energy costs, which is caused by multiple spikes, limit their use in mobile and embedded applications. And recent works have shown that the newly emerged temporal-coding based SNNs converted from DNNs can reduce the computational load effectively. In this paper, we propose a novel method to convert DNNs to temporal-coding SNNs, called TDSNN. Combined with the characteristic of the leaky integrate-andfire (LIF) neural model, we put forward a new coding principle Reverse Coding and design a novel Ticking Neuron mechanism. According to our evaluation, our proposed method achieves 42% total operations reduction on average in large networks comparing with DNNs with no more than 0.5% accuracy loss. The evaluation shows that TDSNN may prove to be one of the key enablers to make the adoption of SNNs widespread.

With the advancements of neural networks, customized accelerators are increasingly adopted in massive AI applications. To gain higher energy efficiency or performance, many hardware design optimizations such as near-threshold logic or overclocking can be utilized. In these cases, computing errors may happen and the computing errors are difficult to be captured by conventional training on general purposed processors (GPPs). Applying the offline trained neural network models to the accelerators with errors directly may lead to considerable prediction accuracy loss. To address this problem, we explore the resilience of neural network models and relax the accelerator design constraints to enable aggressive design options. First of all, we propose to train the neural network models using the accelerators' forward computing results such that the models can learn both the data and the computing errors. In addition, we observe that some of the neural network layers are more sensitive to the computing errors. With this observation, we schedule the most sensitive layer to the attached GPP to reduce the negative influence of the computing errors. According to the experiments, the neural network models obtained from the proposed training outperform the original models significantly when the CNN accelerators are affected by computing errors.

Edge computing is a continuum that includes the computing resources from cloud to things. Ecosystem of things (EoT) is a subsystem of the ecosystem of edge computing, which potentially contains trillions of devices of things and directly interacts with the physical world. This paper surveys the state of the art of EoT by focusing on the computing infrastructure aspect with a forward-looking perspective. We point out a trend of smart edge computing with four types of smartness and intelligence. We address three fundamental questions. 1) What capabilities and how much energy efficiency are the hardware providing? What is the future growth potential? 2) What abstractions are provided by the system software? Are they adequate to support smart edge computing? 3) What ecosystem architectures have been proposed for the coordination of things, the edge, and the cloud? Are they meeting the needs to encourage innovation but avoid unnecessary ecosystem fragmentation? We examine advances from both industry and academia, including research results, visions, and project concepts. We also point out future research directions.

Neural networks (NNs) have been demonstrated to be useful in a broad range of applications, such as image recognition, automatic translation, and advertisement recommendation. State-of-the-art NNs are known to be both computationally and memory intensive, due to the ever-increasing deep structure, i.e., multiple layers with massive neurons and connections (i.e., synapses). Sparse NNs have emerged as an effective solution to reduce the amount of computation and memory required. Though existing NN accelerators are able to efficiently process dense and regular networks, they cannot benefit from the reduction of synaptic weights. In this paper, we propose a novel accelerator, Cambricon-X, to exploit the sparsity and irregularity of NN models for increased efficiency. The proposed accelerator features a processing element (PE)-based architecture consisting of multiple PEs. An indexing module efficiently selects and transfers needed neurons to connected PEs with reduced bandwidth requirement, while each PE stores irregular and compressed synapses for local computation in an asynchronous fashion. With 16 PEs, our accelerator is able to achieve at most 544 GOP/s in a small form factor (6.38 mm2and 954 mW at 65 nm). Experimental results over a number of representative sparse networks show that our accelerator achieves, on average, $7.23\times$ speedup and $6.43\times$ energy saving against the state-of-the-art NN accelerator. We further investigate possibilities of leveraging activation sparsity and multi-issue controller, which improve the efficiency of Cambricon-X. To ease the burden of programmers, we also propose a high efficient library-based programming environment for our accelerator.

Neural networks (NNs) have been demonstrated to be useful in a broad range of applications such as image recognition, automatic translation and advertisement recommendation. State-of-the-art NNs are known to be both computationally and memory intensive, due to the ever-increasing deep structure, i.e., multiple layers with massive neurons and connections (i.e., synapses). Sparse neural networks have emerged as an effective solution to reduce the amount of computation and memory required. Though existing NN accelerators are able to efficiently process dense and regular networks, they cannot benefit from the reduction of synaptic weights. In this paper, we propose a novel accelerator, Cambricon-X, to exploit the sparsity and irregularity of NN models for increased efficiency. The proposed accelerator features a PE-based architecture consisting of multiple Processing Elements (PE). An Indexing Module (IM) efficiently selects and transfers needed neurons to connected PEs with reduced bandwidth requirement, while each PE stores irregular and compressed synapses for local computation in an asynchronous fashion. With 16 PEs, our accelerator is able to achieve at most 544 GOP/s in a small form factor (6.38 mm2and 954 mW at 65 nm). Experimental results over a number of representative sparse networks show that our accelerator achieves, on average, 7.23x speedup and 6.43x energy saving against the state-of-the-art NN accelerator.

In 3-D-stacked memory chips, the problem of power supply integrity (PSI) is aggravating due to the additional through-silicon-via resistance and the higher current density in 3-D power delivery network. In particular, for the 3-D phase-change memory (PCM) well known for its high-amplitude programming current, IR-drop violation poses a serious threat that enforces a strict guard band of requesting concurrence, and consequently reduces the write throughput. This paper presents the implication of an IR-drop phenomenon in a 3-D PCM cube, and investigates IR-drop's impacts on write management in the PCM. From the obtained SPICE simulation results, we find that the issued writes have to meet the IR-drop constraint to be reliably processed, and then propose a PSI conscious write scheduler to improve the write performance within the constraint of the IR-drops and the power budget in the 3-D PCM cube. First, a Bloom-filter-based method is proposed to avoid the invalid write decisions for the PCM. Second, to support fine-grained write management in the cutting-edge PCM, we develop an inexpensive approach, weighted token assignment (WTA), to filter out PSI-unsafe write decisions by employing a support vector machine-based learning model. Last, a write reordering policy is proposed to cooperate with WTA and optimize the total write throughput for better memory performance. In the simulated hybrid main memory composed of both dynamic random access memory and 3-D PCM, the proposed scheduler significantly improves the write throughput.

Emerging highly-parallel and big data applications have renewed the research interest in Processing-in-Memory (PIM) architectures. However, moving powerful processing unit into the CMOS-incompatible DRAM chips is not cost-effective for large capacity memory. In this work, we observe that Non-Volatile Memory is often naturally incorporated with basic logics like Data Comparison Write or Flip-n-Write modules that are essential for cell SET/REST operation. In contrast to other conventional PIM or Near Data Computing (NDC) architectures, ProPRAM, as a typical Active NVM, abandons the design approach of moving accelerators or customized processors into memory devices, but begins with exploiting the existing resources inside the memory chips to accelerate the key non-compute-intensive functions for emerging big data applications. With slight hardware and architectural modification, we succeed to expose the transparent peripheral logics to the application layer through instruction set extension and exploit them for in-field bulk data processing with limited hardware cost. Compared to conventional CPU-centric systems, ProPRAM achieves an excellent optimization on energy-efficiency (15x) for important data-intensive micro-benchmarks and kernels.

A vast array of devices, ranging from industrial robots to self-driven cars or smartphones, require increasingly sophisticated processing of real-world input data (image, voice, radio, ...). Interestingly, hardware neural network accelerators are emerging again as attractive candidate architectures for such tasks. The neural network algorithms considered come from two, largely separate, domains: machine-learning and neuroscience. These neural networks have very different characteristics, so it is unclear which approach should be favored for hardware implementation. Yet, few studies compare them from a hardware perspective. We implement both types of networks down to the layout, and we compare the relative merit of each approach in terms of energy, speed, area cost, accuracy and functionality.

The confluence of 3-D integration and network-on-chip (NoC) provides an effective solution to the scalability problem of on-chip interconnects. In 3-D integration, through-silicon via (TSV) is considered to be the most promising bonding technology. However, TSVs are also precious link resources because they consume significant chip area and possibly lead to routing congestion in the physical design stage. In addition, TSVs suffer from serious yield losses that shrink the effective TSV density. Thus, it is necessary to implement a TSV-economical 3-D NoC architecture in cost-effective design. For symmetric 3-D mesh NoCs, we observe that the TSVs bandwidth utilization is low and they rarely become the contention spots in networks as planar links. Based on this observation, we propose the TSV sharing (TS) scheme to save TSVs in 3-D NoC by enabling neighboring routers to share the vertical channels in a time division multiplexing way. We also investigate different TS implementation alternatives and show how TS improves TSV-effectiveness (TE) in multicore processors through a design space exploration. In experiments, we comprehensively evaluate TSs influence on all layers of system. It is shown that the proposed method significantly promotes TE with negligible performance overhead.

In chip multiprocessors (CMPs), nonuniform cache architecture (NUCA) is often employed to organize last-level cache (LLC) banks through network-on-chip (NoC). Because of the shrinking feature size and unstable operating environment, severe reliability problems unavoidably emerge and cause frequent on-chip component (e.g., cores, cache banks, routers) failures. Typical fault-tolerant CMPs should possess the feature of graceful degradation and function normally with deactivated tiles. However, for CMPs adopting static NUCA, certain physical address areas will become inaccessible when cache banks in a CMP node are isolated from the system. To protect cache from such threats induced by either online or offline faults, we survey several potential solutions and propose the utility-driven node remapping technique by reusing the resources in NoC. In our NoC-assisted remapping scheme, cache accesses to isolated banks are so redirected that cache space contention are successfully balanced and relieved in shared-LLC, thus ensuring the least performance penalty caused by fault isolation. Our experimental results show significant performance improvement over conventional resizing approaches such as set reduction.

3-D technology that stacks silicon dies with through silicon vias (TSVs) is a promising solution to overcome the interconnect scaling problem in giga-scale integrated circuits (ICs). Thermal dissipation is a major challenge for 3-D integration and prior thermal-balanced task scheduling methods for 3-D multiprocessor system-on-chips (MPSoCs) typically balance power gradient across vertical stacks based on the assumption of strong thermal correlation among processing cores within a stack. On the other hand, 3-D MPSoCs typically employ network-on-chip (NoC) as the communication infrastructure which consumes a large portion of the energy budget. As TSVs consume much less energy than horizontal links in 3-D MPSoCs when transmitting the same amount data due to the reduced interconnect distance between vertical adjacent cores, it motivates to allocate heavily communicating tasks within the same vertical stack as much as possible, and thus traffic is restricted in the third dimension to reduce interconnect energy. However, aggregating active tasks within the same stack probably exacerbates the power density and result in hot spots. In this paper, we explore the tradeoff between thermal and interconnect energy when allocating tasks in 3-D Homogeneous MPSoCs, and propose an efficient heuristic. Experimental results show that the proposed technique can reduce interconnect energy by more than 25% on average with almost the same peak temperature when compared with prior thermal-balanced solutions.

Homogeneous manycore processors are emerging in broad application areas, including those with timing requirements, such as real-time and embedded applications. Typically, these processors employ Network-on-Chip (NoC) as the communication infrastructure and core-level redundancy is often used as an effective approach to improve the yield of manycore chips. For a given application's task graph and a task to core mapping strategy, the traffic pattern on the NoC is known a priori. However, when defective cores are replaced by redundant ones, the NoC topology changes. As a result, a fine-tuned program based on timing parameters given by one topology may not meet the expected timing behavior under the new one. To address this issue, a timing similarity metric is introduced to evaluate timing resemblances between different NoC topologies. Based on this metric, a Hungarian method based algorithm is developed to reconfigure a defect-tolerant manycore platform and form a unified application specific virtual core topology of which the timing variations caused by such reconfiguration are minimized. Our case studies indicate that the proposed metric is able to accurately measure the timing differences between different NoC topologies. The standard deviation between the calculated difference using the metric and the difference obtained through simulation is less than 6.58%. Our case studies also indicate that the developed Hungarian method based algorithm using the metric performs close to the optimal solution in comparison to random defect-redundant core assignments.

Concurrent trace is an emerging challenge when debugging multicore systems. In concurrent trace, trace buffer becomes a bottleneck since all trace sources try to access it simultaneously. In addition, the on-chip interconnection fabric is extremely high hardware cost for the distributed trace signals. In this paper, we propose a clustering-based scheme which implements concurrent trace for debugging Network-on-Chip (NoC) based multicore systems. In the proposed scheme, a unified communication framework eliminates the requirement for interconnection fabric which is only used during debugging. With clustering scheme, multiple concurrent trace sources can access distributed trace buffer via NoC under bandwidth constraint. We evaluate the proposed scheme using Booksim and the results show the effectiveness of the proposed scheme.

Three-dimensional (3D) integration and Network-on-Chip (NoC) are both proposed to tackle the on-chip interconnect scaling problems, and extensive research efforts have been devoted to the design challenges of combining both. Through-silicon via (TSV) is considered to be the most promising technology for 3D integration, however, TSV pads distributed across planar layers occupy significant chip area and result in routing congestions. In addition, the yield of 3D integrated circuits decreased dramatically as the number of TSVs increases. For symmetric 3D mesh NoC, we observe that the TSVs' utilization is pretty low and adjacent routers rarely transmit packets via their vertical channels (i.e. TSVs) at the same time. Based on this observation, we propose a novel TSV squeezing scheme to share TSVs among neighboring router in a time division multiplex mode, which greatly improves the utilization of TSVs. Experimental results show that the proposed method can save significant TSV footprint with negligible performance overhead.

Very large scale integrated circuits typically employ Network-on-Chip (NoC) as the backbone for on-chip communication. As technology advances into the nanometer regime, NoCs become more and more susceptible to permanent faults such as manufacturing defects, device wear-out, which hinder the correct operations of the entire system. Therefore, effective fault-tolerant techniques are essential to improve the reliability of NoCs. Prior work mainly focuses on introducing redundancies, which can't achieve satisfactory reliability and also involve large hardware overhead, especially for data path components. In this paper, we propose fine-grained data path salvaging techniques by splitting data path components, i.e., links, input buffers and crossbar into slices, instead of introducing redundancies. As long as there is one fault-free slice for each component, the router can be functional. Experimental results show that the proposed solution achieves quite high reliability with graceful performance degradation even under high fault rate.

Three dimensional (3D) System-on-Chips (SoCs) that typically employ through-silicon vias (TSVs) as vertical interconnects, emerge as a promising solution to continue Moore's law. Whereas, it also brings challenging problems, one of which is the test wrapper chain design and optimization, especially for circuit-partitioned 3D SoCs in which scan chains can cross among layers. Test time is the primary goal for wrapper chain design, both for 2D and 3D SoCs. The 3D SoC wrapper chain design problem can be converted into the well-studied2D one by projecting wrapper chain components of all layers to one virtual layer. Thereafter, we can leverage 2D optimization algorithms to determine the composition of wrapper chains and thus guarantee minimal testing time for 3D SoCs. One specific thing for circuit-partitioned 3D SoCs is that TSVs are needed to connect cross-layer wrapper structures to form the wrapper chains. As TSVs occupy planar chip area and will aggravate the routing congestion problem, it is necessary to reduce TSVs for test purpose as much as possible. In this work, we observe that by varying the connection orders of wrapper chain components, e.g., scan chains and I/O cells, the TSVs consumed vary significantly. Based on the above, we formulate this problem and propose novel heuristic to tackle it. Experimental results show that the proposed solution can save on average 33.2% amount of TSVs when compared to a prior intuitive method.

The limited write endurance of phase change random access memory (PRAM) is one of the major obstacles for PRAM-based main memory. Wear leveling techniques were proposed to extend its lifetime by balancing writes traffic. Another important concern that need to be considered is endurance variation in PRAM chips. When different PRAM cells have distinct endurance, balanced writes will result in lifetime degradation due to the weakest cells. Instead of balancing writes traffic, in this paper we propose wear rate leveling (WRL), a variant of wear leveling, to balance wear rates (i.e., writes traffic/edudrance) of cells across the PRAM chip. After investigating writing behavior of applications and endurance variation, we propose an architecture-level WRL mechanism. Moreover, there is an important tradeoff between endurance improvement and swapping data volume. To co-optimize endurance and swapping, a novel algorithm, Max Hyper-weight Rematching, is proposed to maximize PRAM lifetime and minimize performance degradation. Experimental results show 19x endurance improvement to prior Wear Leveling.

The emerging nanophotonic technology can avoid the limitation of I/O pin count, and provide abundant memory bandwidth. However, current DRAM organization has mainly been optimized for a higher storage capacity and package pin utilization. The resulted data fetching mechanism is quite inefficient in performance and energy saving, and cannot effectively utilize the abundant optical bandwidth in off-chip communication. This paper inspects the opportunity brought by optical communication, and revisits the DRAM memory architecture considering the technology trend towards multiprocessors. In our FlexMemory design, super-line prefetching is proposed to boost system performance and promote energy efficiency, which leverages the abundant photonic bandwidth to enlarge the effective data fetch size per memory cycle. To further preserve locality and maintain service parallelism for different workloads, page folding technique is employed to achieve adaptive data mapping in photonics-connected DRAM chips via optical wavelengths allocation. By combining both techniques, surplus off-chip bandwidth can be utilized and effectively managed adapting to the workloads intensity. Experimental results show that our FlexMemory achieves considerable improvements in performance and energy efficiency.

Topology virtualization techniques are proposed for NoC-based many-core processors with core-level redundancy to isolate hardware changes caused by on-chip defective cores. Prior work focuses on homogeneous cores with symmetric performance and optimizes on-chip communication only. However, core-to-core performance asymmetry due to manufacturing process variations poses new challenges for constructing virtual topologies. Lower performance cores may scatter over a virtual topology, while operating systems typically allocate tasks to continuous cores. As a result, parallel applications are probably assigned to a region containing many slower cores that become bottlenecks. To tackle the above problem, in this paper we present a novel performance-asymmetry-aware reconfiguration algorithm Bubble-Up based on a new metric called core fragmentation factor (CFF). Bubble-Up can arrange cores with similar performance closer, yet maintaining reasonable hop distances between virtual neighbors, thus accelerating applications with higher degree of parallelism, without changing existing allocation strategies for OS. Experimental results show its effectiveness.

Large scale Chip-Multiprocessors (CMPs) generally employ Network-on-Chip (NoC) to connect the last level cache (LLC), which is generally organized as distributed NUCA (non-uniform cache access) arrays for scalability and efficiency. On the other hand, aggressive technology scaling induces severe reliability problems, causing on-chip components (e.g., cores, cache banks, routers) failure due to manufacture defects or on-line hardware faults. Typical degradable CMPs should possess the ability to work around defects by disabling faulty components. For static NUCA architecture, when cache banks attached to a computing node are disabled, however, certain physical address sections will no longer be accessible. Prior approaches such as sets reduction introduced in Intel Xeon processor 7100 series enable turning off cache banks by masking certain sets bits in physical address1, which greatly wastes cache capacity. In this paper, we propose to tackle the above problem in a finer granularity to restrict the capacity loss in NUCA cache. Cache accesses to isolated nodes are redirected based on the utility-driven address remapping scheme that reduces data blocks conflicts in fault-tolerant shared-LLC. We evaluate our technique using GEMS simulator. Experimental results show that address remapping achieves significant improvement over the conventional cache sizing scheme.

A precomputed core test set contains a large number of don't cares (x's) that can be effectively exploited to improve test data compression (TDC). Extending pattern run-length coding, we present a novel strategy that propagates thex's of a reference pattern to a new reference pattern in such a way that the reference pattern is xor-ed with the pattern to be encoded. Thex-propagating strategy can increase the probability of a reference pattern being coding-compatible with the pattern to be encoded, and its validity can be established by filling somex's of the already encoded patterns in backtracing way. How our strategy is used for TDC is demonstrated. Experimental results for large ISCAS89 benchmarks show that, compared to the recently proposed schemes, our technique can effectively improve compression and simplify on-chip decoder, and work better when used for core-unified TDC.

Negative bias temperature instability (NBTI) has been a significant reliability concern in current digital circuit design due to its effect of increasing the path delay with time and in turn degrading the circuit performance. NBTI degradation has strong dependence on input pattern and duty cycles. Based on this observation, we propose to apply multiple input vectors to the combination circuit in a non-uniform way during standby mode. Multiple input vectors can enhance the capability to control the circuit nodes, achieve smaller duty cycles to reduce the stress time of gates and thus mitigate static NBTI. A constrained multi-object optimization model is formalized to find the optimal combination of duty cycles for timing-critical paths, which in turn minimizes the increase of path delay. An ATPG-like procedure is then presented to generate the corresponding input vectors. Experimental results demonstrate that the delay increase of timing-critical paths can be mitigated significantly under long time NBTI effect (10-year) by only applying a small number of vectors.

Thread-level redundancy in Chip Multiprocessors(TLR-CMP) is efficient for soft error tolerance. Process variation causes core-to-core (C2C) performance asymmetry across a chip, which should be taken into consideration for application scheduling. In this paper, two types of variations beyond C2C are introduced, i.e., inter-pair and intra-pair variation in TLR-CMP. Intra-pair performance asymmetry can affect the performance of applications differently. Based on the above observation, we firstly formalize the variation aware scheduling in TLR-CMP as a 0-1 programming problem,to maximize the system weighted throughput. An efficient scheduling algorithm, named IntraVarF&AppSen, is then proposed to tackle this problem, which can be proved to be optimal when the number of applications to be scheduled is equal to the number of core pairs. Simulation on a 64-core CMP shows 2.8%-4% improvement in weighted throughput when compared to prior VarF&AppIPC algorithm.

Homogeneous manycore systems are emerging for tera-scale computation and typically utilize Network-on-Chip (NoC) as the communication scheme between embedded cores. Effective defect tolerance techniques are essential to improve the yield of such complex integrated circuits. We propose to achieve fault tolerance by employing redundancy at the core-level instead of at the microarchitecture level. When faulty cores exist on-chip in this architecture, however, the physical topologies of various manufactured chips can be significantly different. How to reconfigure the system with the most effective NoC topology is a relevant research problem. In this paper, we first show that this problem is an instance of a well known NP-complete problem. We then present novel solutions for the above problem, which not only maximize the performance of the on-chip communication scheme, but also provide a unified topology to Operating System and application software running on the processor. Experimental results show the effectiveness of the proposed techniques.

Homogeneous manycore processors are emerging for tera-scale computation. Effective defect tolerance techniques are essential to improve the yield of such complex integrated circuits. In this paper, we propose to achieve fault tolerance by employing redundancy at the core-level instead of at the microarchitecture-level. When faulty cores existing on-chip in this architecture, how to reconfigure the processor with the most effective topology is a relevant research problem. We present novel solutions for this problem, which not only maximize the performance of the manycore processor, but also provide a unified topology to operating system and application software running on the processor. Experimental results show the effectiveness of the proposed techniques.

A novel test data compression coding scheme, block marking and updating coding, is proposed in this paper. Test data in the test set was divided into successive fixed-length vectors, called blocks, and then they were marked according to their compatibility compared with a reference vector. An operation that is similar to difference and a technique that strategically fills in don 't-care bits are combined to increase the probabilities of compatibility or inverse compatibility. It effectively compresses test data and its decompression structure is very simple. Experimental results of ISCAS-89 benchmark circuits show that the scheme is very effective.