Fused-layer CNN accelerators
The DOI matches its Crossref record; 20 of the 20 references it cites were found in a registry or an index.
Abstract
Deep convolutional neural networks (CNNs) are rapidly becoming the dominant approach to computer vision and a major component of many other pervasive machine learning tasks, such as speech recognition, natural language processing, and fraud detection. As a result, accelerators for efficiently evaluating CNNs are rapidly growing in popularity. The conventional approaches to designing such CNN accelerators is to focus on creating accelerators to iteratively process the CNN layers. However, by processing each layer to completion, the accelerator designs must use off-chip memory to store intermediate data between layers, because the intermediate data are too large to fit on chip. In this work, we observe that a previously unexplored dimension exists in the design space of CNN accelerators that focuses on the dataflow across convolutional layers. We find that we are able to fuse the processing of multiple CNN layers by modifying the order in which the input data are brought on chip, enabling caching of intermediate data between the evaluation of adjacent CNN layers. We demonstrate the effectiveness of our approach by constructing a fused-layer CNN accelerator for the first five convolutional layers of the VGGNet-E network and comparing it to the state-of-the-art accelerator implemented on a Xilinx Virtex-7 FPGA. We find that, by using 362KB of on-chip storage, our fused-layer accelerator minimizes off-chip feature map data transfer, reducing the total transfer by 95%, from 77MB down to 3.6MB per image.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“When the number of worker nodes increases from 2 to 4, MoDNN can speedup the DNN computation by 2.17-4.28×[130]”
From Convergence of Edge Computing and Deep Learning: A Comprehensive Survey (2020) · cites this paper as [130]
In the citing paper
“With regard to specific DNN structures, e.g., CNN, a finer grid partitioning can be applied to minimize communication, synchronization, and memory overhead [130].”
From Convergence of Edge Computing and Deep Learning: A Comprehensive Survey (2020) · cites this paper as [130]
Citations
Cites
Very Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 6×
“Table II shows the results of a similar comparison, but for the larger and newer VGGNet-E network [17].”
BackgroundMethodResultInfluentialImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 4×
“(We observe a similar pattern in other common CNN structures, such as AlexNet [8].”
BackgroundMethodInfluentialGoing deeper with convolutions
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 3×
“Large Scale Visual Recognition Challenge (ILSVRC) [15], uses five convolutional layers, while VGG [17] (winner of a portion of ILSVRC 2014) uses up to 17 convolutional layers.”
MethodOptimizing FPGA-based Accelerator Design for Deep Convolutional Neural Networks
Zhang, C., Li, P., Sun, G., Guan, Y., Xiao, B., & Cong, J. (2015) · Proceedings of the 2015 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays · doi:10.1145/2684746.2689060 · Cited by 2,158 · cited 10×
“The computational dominance of the convolutional layers, coupled with the need for larger and deeper networks, has sparked significant interest in the design and optimization of accelerator structures for these layers [19], [10].”
BackgroundMethodResultInfluentialDaDianNao: A Machine-Learning Supercomputer
Chen, Y., Luo, T., Liu, S., Zhang, S., He, L., Wang, J., Li, L., Chen, T., Xu, Z., Sun, N., et al. (2014) · 2014 47th Annual IEEE/ACM International Symposium on Microarchitecture · doi:10.1109/micro.2014.58 · Cited by 1,492 · cited 3×
“As a result, there has been significant interest in developing and adapting hardware accelerators for CNNs [7] such as GPUs [3], FPGAs [13], [19], [10], [1], and ASICs [2].”
BackgroundMethodTorch7: A Matlab-like Environment for Machine Learning
Collobert, R., Kavukcuoglu, K., & Farabet, C. (2011) · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · Cited by 1,186
cuDNN: Efficient Primitives for Deep Learning
Chetlur, S., Woolley, C., Vandermersch, P., Cohen, J., Tran, J. J., Catanzaro, B., & Shelhamer, E. (2014) · arXiv (Cornell University) · doi:10.48550/arxiv.1410.0759 · Cited by 943 · cited 2×
“As a result, there has been significant interest in developing and adapting hardware accelerators for CNNs [7] such as GPUs [3], FPGAs [13], [19], [10], [1], and ASICs [2].”
MethodA reconfigurable fabric for accelerating large-scale datacenter services
Putnam, A., Caulfield, A. M., Chung, E. S., Chiou, D., Constantinides, K., Demme, J., Esmaeilzadeh, H., Fowers, J., Gopal, G. P., Gray, J., et al. (2014) · 2014 ACM/IEEE 41st International Symposium on Computer Architecture (ISCA) · doi:10.1109/isca.2014.6853195 · Cited by 576
NeuFlow: A runtime reconfigurable dataflow processor for vision
Farabet, C., Martini, B., Corda, B., Akselrod, P., Culurciello, E., & LeCun, Y. (2011) · CVPR 2011 WORKSHOPS · doi:10.1109/cvprw.2011.5981829 · Cited by 369 · cited 1×
“In [6], [5], [16], systolic implementations were proposed.”
Cited by
Items 501–600 of 779
Energy-Efficient Dataflow Scheduling of CNN Applications for Vector-SIMD DSP
Kim, W., Lee, S., Yun, I., Lee, C., Lee, K., & Lee, H.-J. (2022) · IEEE Access · doi:10.1109/access.2022.3197206 · Cited by 4
“In [11] and [21], a method was proposed to reuse the overlapped data in adjacent tiles and in adjacent layers, respectively.”
BackgroundMethodResultInfluentialORSAS: An Output Row-Stationary Accelerator for Sparse Neural Networks
Lin, C., Liu, Y., & Shang, D. (2023) · IEEE Access · doi:10.1109/access.2023.3272564 · Cited by 4
NASA-F: FPGA-Oriented Search and Acceleration for Multiplication-Reduced Hybrid Networks
Shi, H., Xu, Y., Wang, Y., Mao, W., & Wang, Z. (2023) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3327965 · Cited by 4
“(iii) An intermediary solution between the reconfigurable architecture and the stream architecture is layer fusion [36].”
BackgroundAccelerating low rank matrix completion on FPGA
Zhou, S., Kannan, R., & Prasanna, V. K. (2017) · International Conference on ReConFigurable Computing and FPGAs (ReConFig) · doi:10.1109/reconfig.2017.8279771 · Cited by 4
“With the increased interest in energy-efficient acceleration [3], [4], [5], [6], FPGA has become an attractive platform to develop accelerators for computation-intensive algorithms [7], [8], [9], [10], [11].”
BackgroundA Reconfigurable Pipelined Architecture for Convolutional Neural Network Acceleration
Xue, C., Cao, S. C., Jiang, R., & Yang, H. (2018) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas.2018.8351425 · Cited by 4
“The work in [9] and [10] introduces pipeline in the FPGA design of CNN network, which however, did not actually fully pipelined both innerlayer and inter-layer in CNNs.”
MethodDesign Space Exploration of Convolution Algorithms to Accelerate CNNs on FPGA
Kala, S., Paul, D., Jose, B. R., & Nalesh, S. (2018) · International Symposium on Embedded Computing and System Design (ISED) · doi:10.1109/ised.2018.8704043 · Cited by 4
On-Chip Instruction Generation for Cross-Layer CNN Accelerator on FPGA
Hu, Y., Liang, S., Yu, J., Wang, Y., & Yang, H. (2019) · IEEE Computer Society Annual Symposium on VLSI · doi:10.1109/isvlsi.2019.00011 · Cited by 4
“Many cross-layer CNN accelerators have been proposed based on the idea of layer fusion [1].”
BackgroundMethodArtificial Neural Networks in Hardware
Zheng, N., & Mazumder, P. (2019) · Wiley eBooks · doi:10.1002/9781119507369.ch3 · Cited by 4
Design Space Exploration of Matrix–Matrix Convolution Operation
Behera, P., & Deb, A. (2021) · Journal of Circuits Systems and Computers · doi:10.1142/s0218126621502881 · Cited by 4
Training of deep learning pipelines on memory-constrained GPUs via segmented fused-tiled execution
Xu, Y., Raje, S., Rountev, A., Sabin, G., Sukumaran-Rajam, A., & Sadayappan, P. (2022) · ACM SIGPLAN International Conference on Compiler Construction (CC) · doi:10.1145/3497776.3517766 · Cited by 4
“3 Fused-Tiled execution Tiling and fusion have been used in the design of accelerators for inference in DNNs [2, 24].”
MethodInfluentialJiang, B., Cheng, X., Tang, S., Ma, X., Gu, Z., Fu, S., Yang, Q. X., & Liu, M. (2022) · IEEE International Parallel and Distributed Processing Symposium (IPDPS) · doi:10.1109/ipdps53621.2022.00118 · Cited by 4
Reducing off-chip memory traffic in deep CNNs using stick buffer cache
Rakanovic, D., Erdeljan, A., Vranjković, V. S., Vukobratović, B. Z., Teodorovic, P., & STRUHARIK, R. J. R. (2017) · Telecommunications Forum (TELFOR) · doi:10.1109/telfor.2017.8249398 · Cited by 4
“Similar CNN dataflow where accelerator evaluates a CNN in a layer-by-layer fashion, based on the network’s natural structure is presented in [6, 9, 10, 11].”
BackgroundDeep Fusion: A Software Scheduling Method for Memory Access Optimization
Zhuang, Y., Peng, S., Chen, X., Zhou, S., Zhi, T., Li, W., & Liu, S. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-30709-7_22 · Cited by 4
Liu, Z., Leng, J., Lu, G., Wang, C., Chen, Q., & Guo, M. (2020) · CCF Transactions on High Performance Computing · doi:10.1007/s42514-020-00044-7 · Cited by 4
“For example, there is an overlapping area in 2D sliding window convolution that will be computed multiple times (Ragan-Kelley et al. 2013; Alwani et al. 2016).”
BackgroundEfficient Synchronization for GPGPU
Liu, J. (2018) · D-Scholarship@Pitt (University of Pittsburgh) · Cited by 4
Booster: An Accelerator for Gradient Boosting Decision TreesPreprint
He, M., Vijaykumar, T. N., & Thottethodi, M. S. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2011.02022 · Cited by 4
“, [4], [5], [13], [14], [26], [35], [48] exist for NNs, the most popular ML method, only a few exist for GB, the second-most popular method (Section II-E).”
Systolic Array-based Accelerator for Structured State-Space ModelsPreprint
Raja, S., Demirkiran, C., Sarkar, A., Popovic, M., & Joshi, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.21394 · Cited by 4
“Following are the key innovations in ProDF compared to Eyeriss [4] and FusedCNN [2].”
Pooling On-the-Go for NoC-Based Convolutional Neural Network Accelerator
Zhu, W., Chen, Y., & Lu, Z. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-78380-7_9 · Cited by 3
He, Z., Wang, Y., Wu, Z., Han, H., Wei, S., Hu, Y., Li, C., Tu, F., & Yin, S. (2025) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2025.3593946 · Cited by 3
“Building on this, existing scheduling strategies study how to schedule the computing order and resource allocation of all layers in a NN, which mainly focus on optimizing layer-sequential (LS) [3], [32], [51] and layer-pipeline (LP) [2], [6], [21], [50] patterns.”
5G Network Edge Intelligence for Smart Operation and Maintenance of Offshore Wind Power
Gao, Y., Yang, L., Zhu, X., Jiang, C., Wang, H., You, S., & Xu, F. (2026) · Sensors · doi:10.3390/s26041390 · Cited by 4
“In UAV inspections, cloud-edge collaboration enables real-time high-definition video analysis at the edge—utilizing model pruning and hardware acceleration [16,17]—which drastically short-ens response times by transmitting only key features.”
Lin, K.-P., Ding, Y.-C., Lin, C.-Y., Chen, Y., & Huang, C.-T. (2023) · IEEE Solid-State Circuits Letters · doi:10.1109/lssc.2023.3343913 · Cited by 3
“Previous accelerators for CI CNNs [3], [4], [5], [6], [7], [8] have employed the layer-fusion inference method, as illustrated in Fig.”
MethodMastering Computer Vision Inference Frameworks
Pochelu, P., & Castro, O. (2024) · ACM/SPEC International Conference on Performance Engineering (ICPE) · doi:10.1145/3629527.3651430 · Cited by 3
“For example, sequences of convolutions can be merged into one single convolution [1].”
BackgroundSuperHCA: A Super-Resolution Accelerator with Sparsity-Aware Heterogeneous Core Architecture
Hu, Z., Zeng, J., Zhao, X., Zhou, L., & Chang, L. (2024) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas58744.2024.10557990 · Cited by 3
“J. Lee et al. provided the ring architecture and used the layer fusion method to reduce external memory access (EMA) [5] [6], where the large on-chip buffer is required.”
MethodReMove:Leveraging Motion Estimation for Computation Reuse in CNN-Based Video Processing
Khodarahmi, M., Modarressi, M., Elahi, A., & Pakdaman, F. (2024) · CPSSI International Symposium on Cyber-Physical Systems (Applications and Theory) (CPSAT) · doi:10.1109/cpsat64082.2024.10745388 · Cited by 3
“4: The example of fusing two convolutional layers [14] III.”
BackgroundMethodInfluentialYue, Z., Song, X., Liu, T., Hu, X., Zhang, R., Du, Z., Li, W., Guo, Q., & Chen, T. (2025) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca61900.2025.00074 · Cited by 3
“This is because the optimization of the regular access behavior of the weight matrix has been fully discussed in the existing neural network accelerators [1], [8]–[10], [18].”
USEFUSE: Uniform stride for enhanced performance in fused layer architecture of deep neural networks
Ibrahim, M. S., Usman, M., & Lee, J. (2025) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2025.103459 · Cited by 3
“In the current work, the parameter Q is selected as referenced in literature [21, 33].”
InfluentialCOMET: A Framework for Modeling Compound Operation Dataflows with Explicit Collectives
Negi, S., Singhal, M., Ankit, A., Bhoja, S., & Roy, K. (2026) · IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) · doi:10.1109/ispass69572.2026.00074 · Cited by 3
“Fused-Layer [1] proposes a framework to optimize the execution of convolutional neural networks (CNNs) by fusing multiple adjacent layers to minimize off-chip memory accesses.”
InfluentialMinimizing Off-Chip Memory Access for CNN Accelerators
Tewari, S., Kumar, A., & Paul, K. (2021) · IEEE Consumer Electronics Magazine · doi:10.1109/mce.2021.3097697 · Cited by 3
Duan, X., Xie, R., & Han, J. S. (2021) · IEEE International Conference on ASIC (ASICON) · doi:10.1109/asicon52560.2021.9620485 · Cited by 3
On energy complexity of fully-connected layers
Šíma, J., Cabessa, J., & Vidnerová, P. (2024) · Neural Networks · doi:10.1016/j.neunet.2024.106419 · Cited by 3
“…provide upper bounds on the energy complexity of CNN layers, which have been compared to the real energy consumptions estimated for the Simba (Shao et al., 2019) and Eyeriss (Chen et al., 2016) architectures by using the Timeloop/Accelergy software tool (Parashar et al., 2019; Wu et al., 2019).”
MethodInfluentialPark, J., Kim, S., Park, W.-H., Song, S., & Yoo, H. (2024) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas58744.2024.10558649 · Cited by 3
“Also, CNN accelerators reduce EMA by utilizing an on-chip cache through layer fusion [20], and depth-first execution [21].”
MethodZIP-CNN: Design Space Exploration for CNN Implementation within a MCU
Garbay, T., Hachicha, K., Dobiáš, P., Pinna, A., Hocine, K., Dron, W., Lusich, P., Khalis, I., & Granado, B. (2024) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3691343 · Cited by 3
Telepathic Headache: Mitigating Cache Side-Channel Attacks on Convolutional Neural Networks
Chabanne, H., Danger, J., Guiga, L., & Kühne, U. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-78372-3_14 · Cited by 3
Yang, R., Shen, J., Wen, M., Cao, Y., & Li, Y. (2021) · IEEE International Conference on High Performance Computing and Communications; IEEE International Conference on Smart City; IEEE International Conference on Data Science and Systems (HPCC/SmartCity/DSS) · doi:10.1109/hpcc-dss-smartcity-dependsys53884.2021.00044 · Cited by 3
Fused-Layer-based DNN Model Parallelism and Partial Computation Offloading
Li, M., Wang, N., Zhou, H., Duan, Y., & Wu, J. (2022) · Globecom · doi:10.1109/globecom48099.2022.10000779 · Cited by 3
“The FL technique was first proposed to enable DNN parallel processing in the multi-core environment [16].”
MethodEnergy Complexity of Fully-Connected Layers
Šíma, J., & Cabessa, J. (2023) · Lecture notes in computer science · doi:10.1007/978-3-031-43085-5_1 · Cited by 3
“For simplicity, we assume that the energy cost is not optimized across multiple CNN layers, as for instance in [1].”
MethodCarballo-Hernández, W., Pelcat, M., & Berry, F. (2023) · Journal of Signal Processing Systems · doi:10.1007/s11265-023-01898-0 · Cited by 3
“As seen in Table 1, this technique, as many other partitioning methods, is inspired by the fused-layer [22].”
MethodHigh Throughput and Low Bandwidth Demand: Accelerating CNN Inference Block-by-block on FPGAs
Chen, Y., & Tanaka, K. (2024) · Euromicro Conference on Digital System Design (DSD) · doi:10.1109/dsd64264.2024.00073 · Cited by 3
“Compared to [16], our accelerator is highly efficient, without the aforementioned overheads.”
BackgroundOptimizing of Convolutional Neural Network Accelerator
Du, W., Wang, Z., & Chen, D. (2018) · InTech eBooks · doi:10.5772/intechopen.75796 · Cited by 3
“In addition, data-reusing [7, 8, 23] is also an optimization direction for CNN accelerator.”
BackgroundMedusa: A Scalable Interconnect for Many-Port DNN Accelerators and Wide DRAM Controller Interfaces
Shen, Y., Ji, T., Ferdman, M., & Milder, P. A. (2018) · International Conference on Field Programmable Logic and Applications (FPL) · doi:10.1109/fpl.2018.00026 · Cited by 3
Self-citationPerformance-driven Programming of Multi-TFLOP Deep Learning Accelerators
Venkataramani, S., Choi, J., Srinivasan, V., Gopalakrishnan, K., & Chang, L. (2019) · IEEE International Symposium on Workload Characterization · doi:10.1109/iiswc47752.2019.9042017 · Cited by 3
Zhang, N., Wei, X. S., Chen, L., & Chen, H. (2019) · IEEE International Conference on Signal, Information and Data Processing (ICSIDP) · doi:10.1109/icsidp47821.2019.9173510 · Cited by 3
“Multiple studies use high-level synthesis (HLS) tools to accelerate the CNN inference on FPGAs [5] [10].”
MethodSelf-citation(Pen-) Ultimate DNN PruningPreprint
Riera, M., Arnau, J.-M., & González, A. M. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1906.02535 · Cited by 3
“Some of the most popular applications of MLPs are speech recognition and machine translation [26], [27], whereas CNNs are commonly used for image classification [28].”
BackgroundKAPLA: Pragmatic Representation and Fast Solving of Scalable NN Accelerator DataflowPreprint
Li, Z., & Gao, M. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2306.15676 · Cited by 3
“Layer fusion [4] could be modeled as special layer types as they require specific architectural support on spatial accelerators.”
BackgroundOctopus: A Heterogeneous In-network Computing Accelerator Enabling Deep Learning for networkPreprint
Wen, D., Li, T., Li, C., Xia, P., Yang, H. J., & Sun, Z. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2308.11312 · Cited by 3
“2 DL accelerators There are plenty works focusing on DL accelerator in CV and NLP cases [4, 5, 17, 26, 47], however, DL accelerators for in-network computing have not been thoroughly investigated.”
BackgroundA Survey of Distributed Learning in Cloud, Mobile, and Edge Settings
Threadgill, M., & Gerstlauer, A. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2405.15079 · Cited by 3
Optimized Spatial Architecture Mapping Flow for Transformer AcceleratorsPreprint
Xu, H., Tahmasebi, F., Qiao, Y., Tian, H., Kwon, H., & Huang, S. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2410.07407 · Cited by 3
“However, most previous works either focus on GPUs [11], [12] or CNNs [6], [16].”
BackgroundJin, H., Yang, J., Liu, Y., Lyu, B., Zhang, K., & Bleier, N. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.08873 · Cited by 3
Performance Modeling and Optimization for Machine Learning Workloads
Lin, Z. (2023) · eScholarship (California Digital Library) · Cited by 3
Fast and Fusiest: An Optimal Fusion-Aware Mapper for Accelerator DesignPreprint
Andrulis, T., Gilbert, M. S., Sze, V., & Emer, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.15166 · Cited by 3
“Multi-Einsum mappers have also been proposed for particular workloads and/or dataflows (commonly only convolutional neural networks) [6], [8], [9], [11].”
Discovering Efficient Fused Layer Configurations for Executing Multi-Workloads on Multi-Core NPUs
Lee, Y., Kim, H., Yu, Y., Cho, M., Seo, J., & Park, Y. (2024) · Design, Automation, and Test in Europe · doi:10.23919/date58400.2024.10546719 · Cited by 2
“Fused-convolution is a method of combining consecutive convolution layers as well as element-wise operations [16].”
MethodZheng, Z., Huang, Y., & Chen, D. (2024) · Multimedia Systems · doi:10.1007/s00530-024-01528-0 · Cited by 2
Chen, Y., Chiu, Y.-T., Tu, H.-J., & Huang, C.-T. (2024) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2024.3488042 · Cited by 2
“Fortunately, the EMA problem in CNNs can be alleviated by layer fusion [13] (also known as depth-first scheduling [14], [15] or block-based inference flow [16]), which makes it possible for CI CNNs to run on edge devices with limited EMA bandwidth.”
InfluentialRuntime Fault Localization in Deep Neural Network Accelerators
Liu, W.-K., Talukdar, J., Tan, B., & Chakrabarty, K. (2025) · ACM Transactions on Design Automation of Electronic Systems · doi:10.1145/3770920 · Cited by 2
Shi, M., Jain, V., Jiang, W., Fang, C., Joseph, A., Dehaene, W., & Verhelst, M. (2025) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2025.3636451 · Cited by 2
“The core principle of layer fusion [42] involves initiating execution of subsequent layers immediately upon partial completion of preceding layers, rather than waiting for complete layer processing.”
SD-Acc: Accelerating Stable Diffusion through Phase-Aware Sampling and Hardware Co-Optimizations
Wang, Z., He, G., & Fan, H. (2026) · IEEE Transactions on Computers · doi:10.1109/tc.2026.3715379 · Cited by 2
“Our general framework also allows for exploring the trade-off between image quality and acceleration.”
Integer Intelligence: A Reproducible Path from Training to FPGA
Shanker, M. C., & Tee Hui, T. E. O. (2026) · Electronics · doi:10.3390/electronics15051117 · Cited by 2
“Off-chip traffic and latency were reduced by near-sensor and fused-layer execution, where feature maps were kept on-chip and the computation was placed close to the image [25,26].”
Seong, C., Lee, S., Kim, B., & Jeon, D. (2026) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2026.3680701 · Cited by 2
“Prior designs mitigated this issue by adopting a fused-layer dataflow [23], [24], where consecutive MLP layers (MLP1-MLP3) and the subsequent max pooling operation are executed for a single center point without intermediate accesses to external memory.”
Ma, S., Bie, Z., & Liang, X. (2023) · International Conference on Integrated Circuits and Microsystems (ICICM) · doi:10.1109/icicm59499.2023.10366042 · Cited by 2
“The computational load in YOLOv3-Tiny is mainly distributed in CNNs [4], [7].”
BackgroundDNNMapper: An Elastic Framework for Mapping DNNs to Multi-die FPGAs
Li, S. X., Zhou, X., Lu, H., & Wang, K. (2024) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas58744.2024.10558120 · Cited by 2
“Meanwhile, researchers have explored multiple optimization methods and novel architectures targeting DNN layer structures and data access patterns [2–4].”
BackgroundA Resource-Adaptive, highly Parallel Hardware Accelerator for CNNs on ZYNQ-7000 SoC
Li, Y., Xu, K., Liu, Y., Liang, J., Yan, Y., & You, H. (2024) · doi:10.1145/3688863.3689574 · Cited by 2
“They proposed a "fusion" data transfer mode that groups multiple convolution layers into a single module, maximizing on-chip memory usage and reducing off-chip data transfers [1].”
Verma, P., Gupta, T., Das, P., Nali, A. R., Hunsigida, V., & Acharyya, A. (2024) · IEEE International System-on-Chip Conference (SOCC) · doi:10.1109/socc62300.2024.10737722 · Cited by 2
“Reduced CNN models can be obtained through various techniques including model compression [6], quantization [7], weight pruning [8], convolutional layer fusion [9], etc.”
BackgroundMethodDesign and Research of AI Accelerators for Onboard Computing Platforms Targeting YOLOs Network
Wang, Y., Zhou, H., Liu, X., Li, B., & Ge, Y. (2025) · International Conference on Computer Research and Development · doi:10.1109/iccrd64588.2025.10962911 · Cited by 2
“Qiu Jiantao from Tsinghua University [7] and Alwani from Stony Brook University [8] proposed dividing feature maps into smaller tasks to alleviate on-chip memory pressure.”
Wang, S., Wang, L., Li, S., Tian, S., Guo, S., Kang, Z., Zhang, S., & Xu, W. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-30709-7_6 · Cited by 2
In-circuit tuning of deep learning designs
Que, Z., Noronha, D. H., Zhao, R., Niu, X., Wilton, S. J. E., & Luk, W. W. (2021) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2021.102198 · Cited by 2
“To address this issue, the standard convolutions can be fused with a uniform kernel [14, 15], and various convolution types can be fused automatically [5].”
BackgroundA Survey on System-Level Design of Neural Network Accelerators
Seto, K. (2021) · Journal of Integrated Circuits and Systems · doi:10.29292/jics.v16i2.505 · Cited by 2
“17 (a) shows a streaming architecture [1][36][37][38][39][40][41][42] (also called layer-wise pipeline architecture [43], dataflow architecture [34], etc.”
BackgroundPre-RTL DNN Hardware Evaluator With Fused Layer Support
Yang, C.-C., & Chang, T. (2021) · International SoC Design Conference · doi:10.1109/isocc53507.2021.9614027 · Cited by 2
“On the other hand, the layerfusion operation [4][5] fuses multiple layer executions without external intermediate data access, and only requires loading the data from DRAM for the first layer input and storing the data to DRAM for the last layer output in one fusion group.”
MethodYang, C., Hou, J., Wang, Y., Zhang, H., Wang, X., & Geng, L. (2022) · Journal of Circuits Systems and Computers · doi:10.1142/s0218126622502899 · Cited by 2
Gao, Z., Zhang, Y., & Sun, W. (2022) · IEEE/CIC International Conference on Communications in China (ICCC) · doi:10.1109/icccworkshops55477.2022.9896696 · Cited by 2
“Model compression techniques have been proposed to simply the models, such as neural pruning [9][10] and weight quantization [11][12], and combined approaches [13][14].”
MethodMercado, K., Bavikadi, S., & Sai Manoj, P. D. (2023) · Conference on Information Sciences and Systems · doi:10.1109/ciss56502.2023.10089735 · Cited by 2
“Compared to [6], the proposed architecture employs a slightly higher number of LUTs, but a significantly lower number of DSPs.”
BackgroundInfluentialOpTC – A Toolchain for Deployment of Neural Networks on AURIX TC3xx Microcontrollers
Heidorn, C. J. A., Hannig, F., Riedelbauch, D., Strohmeyer, C., & Teich, J. (2024) · Proceedings · doi:10.1007/978-3-658-45018-2_4 · Cited by 2
“In addition, optimizations are performed merely at the layer level of a neural network, which misses the potential of globally optimizing the overall neural network graph, e.g., by layer fusion [19, 20].”
BackgroundML-Fusion: Determining Memory Levels for Data Reuse Between DNN Layers
Zhou, Z., Duan, X., Chen, K., Chen, Y., & Han, J. S. (2024) · ACM Great Lakes Symposium on VLSI (GLSVLSI) · doi:10.1145/3649476.3658698 · Cited by 2
“Fig.1(b) represents the traditional data flow considering layer fusion [3, 5, 7, 10].”
BackgroundBetter Scalability: Improvement of Block-Based CNN Accelerator for FPGAs
Chen, Y., & Tanaka, K. (2024) · IEEE Access · doi:10.1109/access.2024.3514325 · Cited by 2
On-Sensor Convolutional Neural Networks with Early-Exits
Shalby, H. H. Y., Vecchi, A. D., Scandelli, A., Bartoli, P., Trojaniello, D., Roveri, M., & Villa, F. (2025) · doi:10.1109/ciescompanion65073.2025.11010905 · Cited by 2
“Specifically, we present three key strategies for achieving this goal: 1) Depth-First convolutions : The convolutions within the ISPU are computed using an incremental approach (similar to those explored in the works of [11], [12], and [13]).”
Energy-based Optimization for Resource Limited Neural Network Accelerators with Fused-layer Support
Friedrich, S., Wittig, R., Matúš, E., & Fettweis, G. P. (2021) · RENATE · doi:10.34657/36966 · Cited by 2
ACCDSE: A Design Space Exploration Framework for Convolutional Neural Network Accelerator
Li, Z., Wang, L., Dou, Q., Tang, Y., Guo, S., Zhou, H., & Lu, W. (2018) · Communications in computer and information science · doi:10.1007/978-981-10-7844-6_3 · Cited by 2
A Power-Efficient Hybrid Architecture Design for Image Recognition Using CNNs
Choi, J., Srinivasa, S., Tanabe, Y., Sampson, J., & Narayanan, V. (2018) · IEEE Computer Society Annual Symposium on VLSI · doi:10.1109/isvlsi.2018.00015 · Cited by 2
“[11] propose a cache mechanism for keeping intermediate feature values to preserve inter-layer data locality.”
BackgroundMethodTowards In-Circuit Tuning of Deep Learning Designs
Que, Z., Noronha, D. H., Zhao, R., Wilton, S. J. E., & Luk, W. W. (2019) · doi:10.1109/iccad45719.2019.8942117 · Cited by 2
“To efficiently process convolution blocks, InTune generates FPGA designs that fuse their inner convolution layers [5, 13] and inserts different debug instruments for debug and tuning.”
BackgroundMethodTowards memory-efficient processing-in-memory architecture for convolutional neural networks
Wang, Y., Zhang, M., & Yang, J. (2017) · ACM SIGPLAN Notices · doi:10.1145/3140582.3081032 · Cited by 2
High Power-Efficient and Performance-Density FPGA Accelerator for CNN-Based Object Detection
Zhang, G., Zhang, C., Wang, F., Tang, F., Wu, Y., Yang, X., & Liu, Y. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-88004-0_10 · Cited by 2
An SoC Integration Ready VLIW-Driven CNN Accelerator with High Utilization and Scalability
Hu, C.-H., Tseng, I.-H., Kuo, P.-H., & Huang, J.-D. (2022) · IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) · doi:10.1109/aicas54282.2022.9870010 · Cited by 2
AI at the Edge: Challenges, Applications, and Directions
Joshi, D., Desai, N. V., Chowdhury, S. P., Lee, W., Bathen, L. A. D., Wang, S., & Verma, D. C. (2022) · Wiley eBooks · doi:10.1002/9781119892199.ch9 · Cited by 2
AI Accelerators for Standalone Computer
Kim, T., Lee, J., Jung, H., & Kim, S. (2023) · Springer eBooks · doi:10.1007/978-3-031-22170-5_2 · Cited by 2
Zynq SoC-based IPs for Deep Learning Neural Networks
Pranaav Jothi, M., Manoj Kumar, B., Subalakshmi Sarvani, S., & Rani, S. P. J. V. (2023) · Winter Summit on Smart Computing and Networks (WiSSCoN) · doi:10.1109/wisscon56857.2023.10133848 · Cited by 2
“Another proposal was to decrease the number of off-chip accesses by optimization of the data between adjacent layers in the network [6] [9].”
BackgroundEnergy Complexity Model for Convolutional Neural Networks
Šíma, J., Vidnerová, P., & Mrázek, V. (2023) · Lecture notes in computer science · doi:10.1007/978-3-031-44204-9_16 · Cited by 2
“in [1], which means energy complexity (4) is defined as a simple sum of energy costs over separate convolutional and fully-connected layers in N only, while the less energy-intensive max pooling layers are omitted:”
BackgroundWireless Chip-Scale Communications for Neural Network Accelerators
Liñán, R. (2019) · UPCommons institutional repository (Universitat Politècnica de Catalunya) · Cited by 2
Low-power accelerators for cognitive computing
Riera, M. (2020) · UPCommons institutional repository (Universitat Politècnica de Catalunya) · doi:10.5821/dissertation-2117-330738 · Cited by 2
Self-Adaptive Reconfigurable Arrays (SARA): Using ML to Assist Scaling GEMM AccelerationPreprint
Samajdar, A., Pellauer, M., & Krishna, T. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2101.04799 · Cited by 2
“[37] Eyeriss [4] Alwani et al. [1] NeuroCube [16] MAERI [20] TPU [13] Flexflow [22] Tetris [8] Brainwave [7] Simba [35] Tangram [9] Cascades [31] Sigma [28] Planaria [10] SAGAR (This work) Dataflow and Accelerator Design Space Search.”
MethodBarrier-Free Large-Scale Sparse Tensor Accelerator (BARISTA) For Convolutional Neural NetworksPreprint
Gondimalla, A., Gundabolu, S. C., Vijaykumar, T. N., & Thottethodi, M. S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2104.08734 · Cited by 2
“There are many optimizations for dense architectures, focusing on compute [17,19,26,28,32,41], memory [9,16,33], and reuse [5,8].”
BackgroundSlice-and-Forge: Making Better Use of Caches for Graph Convolutional Network AcceleratorsPreprint
Yoo, M., Song, J., Lee, H., Lee, J., Kim, N., Kim, Y., & Lee, J. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2301.09813 · Cited by 2
“DNNs often include of multiple layers, which can be utilized to optimize memory usage [4, 30, 36, 51].”
BackgroundInfluentialExploiting Inter-Operation Data Reuse in Scientific Applications using GOGETAPreprint
Garg, R., Pellauer, M., Rajamanickam, S., & Krishna, T. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2303.11499 · Cited by 2
“…[25], Timeloop [32], CoSA [18], GAMMA [21], Intersteller [43], TPUv1 [20], MAERI [26], Eyeriss [7], Eyeriss v2 [8] FusedCNN [3], FLAT [22], FlashAttention [9], X-layer [39], ISOSceles [44], TileFlow [45], OMEGA [14] Can reuse among adjacent ops only when there is no delayed dependency SET [6],…”
InfluentialChen, Y.-C., Ando, S., Fujiki, D., Takamaeda-Yamazaki, S., & Yoshioka, K. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2312.06086 · Cited by 2
“A similar commonly-used technique is tiling [1, 11], which, in the context of CL computing, involves blockwise traversal of the feature map to reduce the maximum data size required at any given time, thereby diminishing storage requirements[4, 5].”
BackgroundMethodMulti-Way Data Representation: A Comprehensive Survey on Tensor Decomposition in Machine LearningPreprint
Tatyana, A. (2025) · Engineering Archive (engrXiv) · doi:10.31224/4421 · Cited by 2
Gilbert, M., Andrulis, T., Sze, V., & Emer, J. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2602.15172 · Cited by 2
“In such cases, a loop over p directly under a storage node for A allows only a line buffer to be held in storage for A , which reduces memory requirements [1], [5], [7], [11], [18].”
A Computational Efficient Architecture for Extremely Sparse Stereo Network
Huang, T., Wu, S.-S., Klopp, J. P., Yu, P.-H., & Chen, L. (2021) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas51556.2021.9401565 · Cited by 1
AMAIX In-Depth: A Generic Analytical Model for Deep Learning Accelerators
Zurstraßen, N., Jünger, L., Kogel, T., Keding, H., & Leupers, R. (2022) · International Journal of Parallel Programming · doi:10.1007/s10766-022-00728-3 · Cited by 1
“Both methods have been applied in DLA design [3,5,6,14,12].”
MethodYe, J., Yanagisawa, M., & Shi, Y. (2022) · Electronics · doi:10.3390/electronics11152356 · Cited by 1
“As a result, as with layer fusion [12], significantly large on-chip memory (in several MBs) is required, which limits the utilization of these inter-layer reuse-based methods in low-cost memory-constrained designs.”
BackgroundMethodInfluentialA Fast Compressed Hardware Architecture for Deep Neural Networks
Ansari, A., Shelton, A., Ogunfunmi, T., & Panchbhaiyye, V. (2022) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas48785.2022.9937651 · Cited by 1
Jassim, S. A., & Eltahir, I. K. (2024) · Bulletin of Electrical Engineering and Informatics · doi:10.11591/eei.v13i6.7683 · Cited by 1
An Analysis of Network Overhead in Distributed TinyML
Hollingsworth, K., Nian, S., Gutierrez, A., & Padmanabhan, A. (2024) · IEEE/ACM Symposium on Edge Computing (SEC) · doi:10.1109/sec62691.2024.00051 · Cited by 1
Metadata
- Authors
- Alwani, M., Chen, H., Ferdman, M., & Milder, P.
- Affiliations
- Stony Brook University (US)
- Venue
- 2016 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO)
- Publisher
- IEEE
- Published
- 2016-10
- Pages
- 1-12
- ISSN
- 1072-4451 (ISSN-L)
- OpenAlex
- W4249932213
- Cited by
- 662
- References
- 20
- Citations by year
- 2026: 20 · 2025: 51 · 2024: 72 · 2023: 67 · 2022: 87 · 2021: 103 · 2020: 72 · 2019: 90 · 2018: 76 · 2017: 23 · 2016: 1
- FWCI
- 21.98
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE/ACM International Symposium on Microarchitecture (MICRO) (publishedVersion)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Hardware and Architecture › Embedded Systems Design Techniques
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Keywords
- CNN accelerator, layer fusion, data flow design, FPGA implementation, convolutional neural network inference
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.