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 401–500 of 779
NASA: Neural Architecture Search and Acceleration for Hardware Inspired Hybrid Networks
Shi, H., You, H., Zhao, Y. K., Wang, Z., & Lin, Y. (2022) · International Conference on Computer Aided Design · doi:10.1145/3508352.3549478 · Cited by 8
PipeFuser: Building Flexible Pipeline Architecture for DNN Accelerators via Layer Fusion
Zhou, X., Li, S., Lu, H., & Wang, K. (2024) · Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/asp-dac58780.2024.10473790 · Cited by 8
“Additionally, the layer fusion technique has also been utilized in some pipeline accelerators to reduce the on-chip memory requirement [2], [11].”
MethodSimuNN: A Pre-RTL Inference, Simulation and Evaluation Framework for Neural Networks
Cao, S. C., Deng, W., Bao, Z., Xue, C., Xu, S., & Zhang, S. (2020) · IEEE Journal on Emerging and Selected Topics in Circuits and Systems · doi:10.1109/jetcas.2020.2993854 · Cited by 8
“Recently, several FPGA [22]–[25] and ASIC [26]–[30] based accelerators are proposed to speed up the inference of neural networks in the embedded user end.”
BackgroundBaek, E., Lee, E., Kang, T., & Kim, J. (2022) · IEEE Transactions on Computers · doi:10.1109/tc.2022.3218428 · Cited by 8
“W ITH the soaring popularity of neural networks (NNs), many researchers have put a lot of effort into designing NN accelerators targeting a single NN [1], [2], [3], [4], [5].”
BackgroundMethodGenetic Algorithm Augmented Inception-Net based Image Classifier Accelerated on FPGA
Kaziha, O., Bonny, T., & Jarndal, A. (2023) · Multimedia Tools and Applications · doi:10.1007/s11042-023-15178-3 · Cited by 9
Lightweight Instruction Set for Flexible Dilated Convolutions and Mixed-Precision Operands
Friedrich, S., Sampath, S. B., Wittig, R., Vemparala, M. R., Fasfous, N., Matúš, E., Stechele, W., & Fettweis, G. P. (2023) · International Symposium on Quality Electronic Design (ISQED) · doi:10.1109/isqed57927.2023.10129341 · Cited by 8
“Note that we also include 96 kB for the program of a RISC-V control processor to support a flexible number of fused-layers [18], [24].”
MethodImplementation of Fully-Pipelined CNN Inference Accelerator on FPGA and HBM2 Platform
Nguyen, V.-C., & NAKASHIMA, Y. (2023) · IEICE Transactions on Information and Systems · doi:10.1587/transinf.2022edp7155 · Cited by 8
Architecture-Aware Optimization of Layer Fusion for Latency-Optimal CNN Inference
Yoon, M., & Choi, J. (2023) · IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) · doi:10.1109/aicas57966.2023.10168659 · Cited by 8
“[1] assessed layer fusion strategies but faced scalability issues and only considered simple DNN models.”
BackgroundMethodInfluentialTowards Out-of-core Neural Networks on Microcontrollers
Miao, H., & Lin, F. X. (2022) · IEEE/ACM Symposium on Edge Computing (SEC) · doi:10.1109/sec54971.2022.00008 · Cited by 7
“A case for out-of-core NNs Can an MCU execute NNs that far exceed its physical memory size? A proven wisdom is to dynamically swap tiles of NN layers between memory tiers [19].”
BackgroundEnabling Multiple Tensor-wise Operator Fusion for Transformer Models on Spatial Accelerators
Xu, L., Mo, Z., Wang, Q., Jiang, J., & Jing, N. (2024) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3649329.3657317 · Cited by 7
“Some other studies proposes other fixed dataflow combinations on fusing two operators, e.g. OS-OS [1, 10].”
Adyna: Accelerating Dynamic Neural Networks with Adaptive Scheduling
Li, Z., Yang, B., Li, J., Chen, T., Li, X., & Gao, M. (2025) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca61900.2025.00049 · Cited by 7
“There are mainly three candidate architectures with such spatial parallelism support, namely multiplexable GPUs [43], [45], [60], multi-tenant DNN accelerators (M-tenant) [22], [33], and multi-tile DNN accelerators (M-tile) [2], [6], [18], [69], as listed in Table II.”
DIF-LUT Pro: An Automated Tool for Simple yet Scalable Approximation of Nonlinear Activation on FPGA
Liu, Y., Li, S., Li, Y., Chen, R., Li, S., Yu, J. W., & Wang, K. (2025) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2025.3576333 · Cited by 7
“The linear operations, such as convolution and matrix multiplication, dominate most of the computational overhead in deep learning algorithms and have been extensively studied in recent years to enhance the performance of accelerators [4], [5], [6].”
ALF: Autoencoder-based Low-rank Filter-sharing for Efficient Convolutional Neural Networks
Frickenstein, A., Vemparala, M. R., Fasfous, N., Hauenschild, L., Nagaraja, N.-S., Unger, C., & Stechele, W. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218501 · Cited by 7
“Practically, such codependent layers can be fused with some advanced scheduling techniques [28], eliminating this overhead.”
BackgroundACNPU: A 4.75TOPS/W 1080P@30FPS Super Resolution Accelerator With Decoupled Asymmetric Convolution
Yang, T.-H., & Chang, T. (2023) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3336468 · Cited by 7
“This is achieved through a holistic model fusion strategy rather than a localized layer fusion as discussed in [16].”
BackgroundImproving Data Reuse in NPU On-chip Memory with Interleaved Gradient Order for DNN Training
Kim, J.-W., Na, S., Lee, S., Lee, S., & Huh, J. (2023) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3613424.3614299 · Cited by 7
Hu, Z., Zeng, J., Zhao, X., Zhou, L., & Chang, L. (2024) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2024.3425753 · Cited by 7
“The above three works are baseline [15], Zhao et al. [14], Lee et al. [5]. and 0.06dB drop in PSNR, respectively, demonstrating that anchor-based residual learning can improve the quantization performance.”
RRAM Based Buffer Design for Energy Efficient CNN Accelerator
Guo, K., Yu, J., Ning, X., Hu, Y., Wang, Y., & Yang, H. (2018) · IEEE Computer Society Annual Symposium on VLSI · doi:10.1109/isvlsi.2018.00085 · Cited by 7
“As suggested by [13], when the network is processed layer by layer, the result of one layer needs to be written back to external memory it is larger than the output buffer.”
BackgroundEnergy-Efficient Design of Advanced Machine Learning Hardware
Hanif, M. A., Hafiz, R., Javed, M. U., Rehman, S., & Shafique, M. (2019) · Springer eBooks · doi:10.1007/978-3-030-04666-8_21 · Cited by 7
DC-CNN: Computational Flow Redefinition for Efficient CNN through Structural Decoupling
Yu, F., Qin, Z., Wang, D., Xu, P., Liu, C., Tian, Z., & Chen, X. '. (2020) · Design, Automation, and Test in Europe · doi:10.23919/date48585.2020.9116429 · Cited by 7
“Many previsou works have been targeting at optmizing the computation flow of convolutional neural networks [6], [9].”
BackgroundA Novel CONV Acceleration Strategy Based on Logical PE Set Segmentation for Row Stationary Dataflow
Zhang, B., Gu, H., Wang, K., & Yang, Y. (2021) · IEEE Transactions on Computers · doi:10.1109/tc.2021.3089366 · Cited by 7
NEMO-CNN: An Efficient Near-Memory Accelerator for Convolutional Neural Networks
Brown, G., Tenace, V., & Gaillardon, P. (2021) · IEEE International Conference on Application-specific Systems, Architectures and Processors (ASAP) · doi:10.1109/asap52443.2021.00016 · Cited by 7
On The Efficiency of Sparse-Tiled Tensor Graph Processing For Low Memory Usage
Cipolletta, A., & Calimera, A. (2021) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18074.2021.9586154 · Cited by 7
Accelerating Video Captioning on Heterogeneous System Architectures
Huang, H.-R., Hong, D., Wu, J., Chen, K.-F., Liu, P., & Hsu, W. (2022) · ACM Transactions on Architecture and Code Optimization · doi:10.1145/3527609 · Cited by 8
“Using deep learning accelerators to speed up DNN inference is a feasible solution, making accelerator design a hot research topic in recent years [2, 7, 17, 28, 36, 40, 46, 48].”
BackgroundOnSRAM: Efficient Inter-Node On-Chip Scratchpad Management in Deep Learning Accelerators
Pal, S., Venkataramani, S., Srinivasan, V., & Gopalakrishnan, K. (2022) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3530909 · Cited by 7
“, operation fusion [17] and feature computation partitioning [7, 58].”
BackgroundANSA: Adaptive Near-Sensor Architecture for Dynamic DNN Processing in Compact Form Factors
Pinkham, R., Erhardt, J., de Salvo, B., Berkovich, A., & Zhang, Z. (2022) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2022.3228725 · Cited by 7
“Methods of reducing the cache capacity have been proposed by modifying the layer processing scheme [6], [7], [8], [9].”
BackgroundMethodEcoFlow: Efficient Convolutional Dataflows on Low-Power Neural Network Accelerators
Orosa, L., Koppula, S., Umuroglu, Y., Kanellopoulos, K., Gómez-Luna, J., Blott, M., Vissers, K., & Mutlu, O. (2023) · IEEE Transactions on Computers · doi:10.1109/tc.2023.3272282 · Cited by 7
Compilation and Optimizations for Efficient Machine Learning on Embedded Systems
Zhang, X., Chen, Y., Hao, C., Huang, S., Li, Y., & Chen, D. (2023) · Springer eBooks · doi:10.1007/978-3-031-39932-9_3 · Cited by 7
Zhou, C., Hassman, Z., Shah, D., Richard, V., & Li, Y. (2024) · ACM SIGPLAN International Conference on Compiler Construction (CC) · doi:10.1145/3640537.3641566 · Cited by 7
“In traditional NCHW alignment, tensors are arranged first by the number of images (batch size, N), then channels (C), followed by height (H), and lastly width (W).”
BackgroundA Tile-based Fused-layer CNN Accelerator for FPGAs
Indirli, F., Erdem, A., & Silvano, C. (2020) · IEEE International Conference on Electronics, Circuits and Systems (ICECS) · doi:10.1109/icecs49266.2020.9294981 · Cited by 7
“This approach was introduced in [5]: this architecture is made by multiple subsequent layers, in which only the first layer loads an input region from the external memory, while the other layers will exchange data internally in the FPGA.”
BackgroundMethodEnergy-Efficient Accelerator Design for Deformable Convolution NetworksPreprint
Xu, D., Chu, C., Liu, C., Wang, Y., Li, H., Li, X., & Cheng, K.-T. T. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.02547 · Cited by 7
“Inspired by the neural network fusion techniques [34], we try to fuse the different processing stages such that the intermediate data can be reused via on-chip buffers without additional external memory accesses.”
MethodSPEC CPU2026: Characterization, Representativeness, and Cross-Suite ComparisonPreprint
Li, R., Jacob, A., Yadwadkar, N. J., & John, L. K. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2605.03713 · Cited by 7
“Recent advancements have leveraged the massive parallelism of GPUs [2, 22, 59, 85, 96, 103, 133], the adaptability of FPGAs [1, 20, 31, 105, 131], and custom domain-specific designs [21, 23, 30, 45, 46, 51, 52, 126] to sustain the performance requirements of next-generation ML applications.”
Li, J., Li, K., Un, K.-F., Yu, W.-H., Martins, R. P., & Mak, P. (2024) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2024.3505299 · Cited by 6
“Thus, we employ large on-chip BRAMs for feature and shortcut [11], [12], [13], [14], [15] to reduce the repetitive amount of DRAM exchange.”
StreamNet++: Memory-Efficient Streaming TinyML Model Compilation on Microcontrollers
Hsu, C.-F., Zheng, H.-S., Liu, Y.-Y., & Yeh, T. T. (2024) · ACM Transactions on Embedded Computing Systems · doi:10.1145/3706107 · Cited by 6
Hsieh, S.-W., Yuan, C.-H., Lin, M.-H., Tsai, P.-Y., Nian, Y.-Y., Cheng, C.-Y., Chih, H.-W., Chiang, P.-H., Chiang, M.-H., Kuo, Y.-J., et al. (2025) · Digest of technical papers/Digest of technical papers - IEEE International Solid-State Circuits Conference · doi:10.1109/isscc49661.2025.10904763 · Cited by 6
LVMScissor: Split and Schedule Large Vision Model Inference on Mobile Edges via Salp Swarm Algorithm
Liu, Y., Lu, R., & Wang, D. (2025) · IEEE Transactions on Mobile Computing · doi:10.1109/tmc.2025.3550519 · Cited by 6
Enhanced YOLO with FPGA hardware acceleration for aluminum sheet defect detection
Xia, F., Nan, G., Jia, Z., & Wang, D. (2025) · Future Generation Computer Systems · doi:10.1016/j.future.2025.108189 · Cited by 6
Gong, Y. D., Huang, L., Chang, H., Liang, R., Yang, C., Tang, Z., Hu, J., & Yuan, B. (2025) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3725843.3756023 · Cited by 6
“Processing : [1][2][ Processing Writing”
Huang, A.-J., Hung, J.-H., & Chang, T. (2023) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2023.3253803 · Cited by 6
“A promising way is to use layer fusion [12] that stores all intermediate feature maps on the chip to eliminate the external memory bandwidth.”
BackgroundMethodInfluentialeNODE: Energy-Efficient and Low-Latency Edge Inference and Training of Neural ODEs
Zhu, J., Tao, Y., & Zhang, Z. (2023) · IEEE International Symposium on High-Performance Computer Architecture (HPCA) · doi:10.1109/hpca56546.2023.10070935 · Cited by 6
“Inspired by the depth-first NN processing [1], [10], [11], we propose transforming a highorder integrator’s compute graph to reduce external DRAM access and minimize the on-chip SRAM size.”
MethodFusing In-storage and Near-storage Acceleration of Convolutional Neural Networks
Okafor, I., Ramanathan, A. K., Challapalle, N. R., Li, Z., & Narayanan, V. (2023) · ACM Journal on Emerging Technologies in Computing Systems · doi:10.1145/3597496 · Cited by 6
“[2] investigate fusing layers to reduce data movement between layer processing.”
BackgroundSong, J., Liang, R., Yuan, B., & Hu, J. (2024) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2024.3429419 · Cited by 6
“The exploration of layer fusions [33] remains an area for future research.”
Effective Runtime Fault Detection for DNN Accelerators
Liu, W.-K., Talukdar, J., Tan, B., & Chakrabarty, K. (2024) · Proceedings - Asian Test Symposium/Proceedings · doi:10.1109/ats64447.2024.10915421 · Cited by 6
“To address the intensive computational demands of DNNs, there is growing interest in specialized hardware accelerators for DNNs to enhance performance and energy efficiency [5]–[9].”
FTConv: FPGA Acceleration for Transposed Convolution Layers in Deep Neural Networks
Tang, Z., Luo, G., & Jiang, M. F. (2019) · ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA) · doi:10.1145/3289602.3293976 · Cited by 6
AMAIX: A Generic Analytical Model for Deep Learning Accelerators
Jünger, L., Zurstraßen, N., Kogel, T., Keding, H., & Leupers, R. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-60939-9_3 · Cited by 6
“Both methods have been applied in DLAs design [2, 4, 5, 10].”
MethodJoint Optimization of Data Transfer and Co-Execution for DNN in Edge Computing
Fu, Z., Zhou, Y., Wu, C., & Zhang, Y. (2021) · International Conference on Communications · doi:10.1109/icc42927.2021.9500513 · Cited by 6
“Fused-layer methods [13] [14] reduce these overheads by removing intermediate transmissions, but they introduce redundant calculations to generate common data both required by the local and edge server.”
MethodSlice-and-Forge: Making Better Use of Caches for Graph Convolutional Network Accelerators
Yoo, M., Song, J., Lee, H., Lee, J., Kim, N., Kim, Y., & Lee, J. (2022) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1145/3559009.3569693 · Cited by 6
“DNNs often include of multiple layers, which can be utilized to optimize memory usage [4, 30, 36, 51].”
BackgroundInfluentialOlympus: Reaching Memory-Optimality on DNN Processors
Cai, X., Wang, Y., Tu, K., Gao, C., & Zhang, L. (2021) · IEEE Transactions on Computers · doi:10.1109/tc.2021.3112262 · Cited by 6
Li, J., Yan, G., Lü, W., Gong, S., Jiang, S., Wu, J., & Li, X. (2018) · ACM Transactions on Design Automation of Electronic Systems · doi:10.1145/3275243 · Cited by 6
“EIE (Han et al. 2016) is an energy-efficient inference engine that performs inference on the compressed FC layers and accelerates the resulting sparse matrix-vector multiplication.”
Ernet Family: Hardware-Oriented Cnn Models For Computational Imaging Using Block-Based Inference
Huang, C.-T. (2020) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp40776.2020.9054521 · Cited by 6
“For example, the layer fusion in [18] proposes a pyramid inference flow and suggests to reuse the overlapped features between moving blocks.”
BackgroundFPGA-based accelerator for convolution operations
Cao, Y., Wei, X., Qiao, T., & Chen, H. (2019) · IEEE International Conference on Signal, Information and Data Processing (ICSIDP) · doi:10.1109/icsidp47821.2019.9172934 · Cited by 6
“As shown in Table IV, our accelerator is better than [5], [11], [12]. and increased by at least 19.3% in terms of performance density.”
BackgroundInfluentialSelf-citationAccelerating convolutional neural networks on FPGAs
Lu, L., Zheng, S., Xiao, Q., Chen, D., & Liang, Y. E. (2019) · Scientia Sinica Informationis · doi:10.1360/n112018-00291 · Cited by 6
“文献 [16] 首次提出 fusion 的技术 , 通过将中间结果一直保留到下一层的计算来减少带宽需求 .”
Compiling Neural Networks for a Computational Memory AcceleratorPreprint
Kourtis, K., Dazzi, M., Ioannou, N., Grosser, T., Sebastian, A., & Eleftheriou, E. S. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.04293 · Cited by 6
“One of the key challenges that recent NN accelerators try to address is how to organize the computation to maximize data reuse within a given memory hierarchy across the different options of reusing input data, reusing weights, and reusing intermediate results [2, 5, 42, 43].”
BackgroundOn the Efficiency of Convolutional Neural NetworksPreprint
Lavin, A. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2404.03617 · Cited by 6
PipeOrgan: Efficient Inter-operation Pipelining with Flexible Spatial Organization and InterconnectsPreprint
Garg, R., Kwon, H., Qin, E., Chen, Y., Krishna, T., & Lai, L. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2405.01736 · Cited by 6
“Therefore, current works are actively investigating inter-layer pipelining or inter-operator (operator means tensor-operator) pipelining in order to reuse the portions of the intermediate feature maps [2], [3], [8], [35], [37], [40], [44].”
BackgroundMethodReconfigurable hardware acceleration of CNNs on FPGA-based smart cameras
Abdelouahab, K. (2018) · theses.fr (ABES) · doi:10.70675/88c8fb7dz3228z4c3cz926cz62f365d6d2d7 · Cited by 6
Lin, K.-P., Liu, J.-H., Liao, H.-C., Wu, J.-Y., Wu, T., & Huang, C.-T. (2025) · IEEE Journal of Solid-State Circuits · doi:10.1109/jssc.2025.3547982 · Cited by 5
“Previous work [14], [15], [16], [17] has focused on high-resolution applications using a layer-fusion (LF) workflow [18] and proposed optimal approaches in the image-spatial domain to reduce the substantial external memory access (EMA) of intermediate FMs.”
Wang, Y., & Zhao, H. (2025) · Electronics · doi:10.3390/electronics14112127 · Cited by 5
In-Memory Computing Architecture for a Convolutional Neural Network Based on Spin Orbit Torque MRAM
Huang, J.-Y., Syu, J.-L., Tsou, Y.-T., Kuo, S., & Chang, C. (2022) · Electronics · doi:10.3390/electronics11081245 · Cited by 5
“The MRAM model used in our research is the same as that presented in [17].”
MethodBae, S.-H., Lee, H., & Kim, H. (2024) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2024.3504513 · Cited by 5
Efficient Mapping of CNNs onto Tightly Coupled Processor Arrays
Heidorn, C. J. A. (2019) · Journal of Computers · doi:10.17706/jcp.14.8.541-556 · Cited by 5
“To decrease the amount of off-chip/on-chip transfers, layer fusion may be applied, which was first proposed by Alwani et al . in [10].”
BackgroundMethodInfluentialCASH: compiler assisted hardware design for improving DRAM energy efficiency in CNN inference
Sarma, A., Jiang, H., Pattnaik, A., Kotra, J., Kandemir, M., & Das, C. R. (2019) · International Symposium on Memory Systems (MEMSYS) · doi:10.1145/3357526.3357536 · Cited by 5
“In Fused Layer CNN [1], computations are staged such that intermediate data is consumed as soon as it is generated in order to avoid costly overflows to the FPGA block RAMs.”
MethodArea-optimal FPGA implementation of the YOLO v2 algorithm using High-Level Synthesis
Sikka, P., Asati, A., & Shekhar, C. (2020) · IEEE Uttar Pradesh Section International Conference on Electrical, Electronics and Computer Engineering (UPCON) · doi:10.1109/upcon50219.2020.9376441 · Cited by 5
Towards Flexible and Compiler-Friendly Layer Fusion for CNNs on Multicore CPUs
Lin, Z., Georganas, E., & Owens, J. D. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-85665-6_15 · Cited by 5
X-Layer: Building Composable Pipelined Dataflows for Low-Rank Convolutions
Vedula, N., Hojabr, R., Khonsari, A., & Shriraman, A. (2021) · International Conference on Parallel Architectures and Compilation Techniques (PACT) · doi:10.1109/pact52795.2021.00015 · Cited by 5
Hardware Platform-Aware Binarized Neural Network Model Optimization
Vo, Q. H., Asim, F., Alimkhanuly, B., Lee, S., & Kim, L.-W. (2022) · Applied Sciences · doi:10.3390/app12031296 · Cited by 5
“Other well-known methods used for hardware optimization can be found in [24–32].”
MethodA Real-Time 1280 × 720 Object Detection Chip With 585 MB/s Memory Traffic
Chang, K.-W., Shih, H.-T., Chang, T., Tsai, S.-H., Yang, C.-C., Wu, C., & Huang, C.-M. (2022) · IEEE Transactions on Very Large Scale Integration (VLSI) Systems · doi:10.1109/tvlsi.2022.3149768 · Cited by 5
“In contrast, layer-fusion DLAs [21]–[23] fuse multiple-layer computation that processes the next layer once its input data are ready.”
BackgroundInfluentialData and Computation Reuse in CNNs Using Memristor TCAMs
de Moura, R. F., de Lima, J. P. C., & Carro, L. (2022) · ACM Transactions on Reconfigurable Technology and Systems · doi:10.1145/3549536 · Cited by 5
“Also, a single�-bit TCAM can represent any �-PLA as long as � ≤ � and the number of entries required is within [1, 2].”
BackgroundInfluentialEnergy Complexity of Convolutional Neural Networks
Šíma, J., Vidnerová, P., & Mrázek, V. (2024) · Neural Computation · doi:10.1162/neco_a_01676 · Cited by 5
“For simplicity, we do not consider the energy optimization across multiple layers (Alwani et al., 2016), which means energy complexity (see equation 3.1) is defined as a simple sum of energy costs only over separate convolutional layers in N (including fully connected layers), while the less…”
BackgroundMethodInfluentialA 16.41 TOPS/W CNN Accelerator with Event-Based Layer Fusion for Real-Time Inference
Wang, J., Li, L., Dai, Z., Jiang, Y., & Cui, X. (2024) · IEEE International Symposium on Circuits and Systems proceedings · doi:10.1109/iscas58744.2024.10558289 · Cited by 5
“This procedure degrades performance and energy efficiency [9].”
BackgroundNi, Y., Chen, W., Cui, W., Zhou, Y., & Qiu, K. (2018) · Design, Automation, and Test in Europe · doi:10.23919/date.2018.8342193 · Cited by 5
“Due to the intrinsic layer dependence of CNNs, each point in a hidden layer of the network only depends on a well-defined region of the previous layer [12].”
BackgroundHuang, P.-T., Wu, I., Lo, C.-Y., & Hwang, W. (2021) · IEEE Open Journal of Circuits and Systems · doi:10.1109/ojcas.2020.3041685 · Cited by 5
“Moreover, the propose TRC can support irregular data accesses, spatial tiling and inter-layer fusion [22] techniques.”
Accelerating Deformable Convolution Networks with Dynamic and Irregular Memory Accesses
Chu, C., Liu, C., Xu, D., Wang, Y., Luo, T., Li, H., & Li, X. (2023) · ACM Transactions on Design Automation of Electronic Systems · doi:10.1145/3597431 · Cited by 5
“Inspired by the neural network fusion techniques [36], we try to fuse the different processing stages such that the intermediate data can be reused via on-chip buffers without additional external memory accesses.”
MethodSanchez, J., Soltani, N., Kulkarni, P., Chamarthi, R. V., & Tabkhi, H. (2018) · Lecture notes in computer science · doi:10.1007/978-3-319-94340-4_4 · Cited by 5
Hardware Efficient Convolution Processing Unit for Deep Neural Networks
Hazarika, A., Poddar, S., & Rahaman, H. (2019) · International Symposium on Devices, Circuits and Systems (ISDCS) · doi:10.1109/isdcs.2019.8719278 · Cited by 5
“Pipeline architecture in FPGA design for CNN acceleration is introduced in [10] and [11].”
MethodSkipping CNN Convolutions Through Efficient Memoization
de Moura, R. F., Santos, P. C., de Lima, J. P. C., Alves, M. A. Z., Beck, A. C. S., & Carro, L. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-27562-4_5 · Cited by 5
“Generally, after a convolution, an activation function is applied on the output feature maps, and, occasionally, it is also followed by a subsampling operation[1].”
BackgroundMethodLowering Dynamic Power of a Stream-based CNN Hardware Accelerator
Piyasena, D., Wickramasinghe, R., Paul, D., Lam, S.-K., & Wu, M. (2019) · IEEE International Workshop on Multimedia Signal Processing (MMSP) · doi:10.1109/mmsp.2019.8901777 · Cited by 5
“In latency driven approaches, streaming accelerators are developed using multiple layer fusion [14], [15], and changing tile granularity [13], with single/multiple FPGA approaches.”
MethodA Tile-based Fused-Layer Approach to Accelerate DCNNs on Low-Density FPGAs
Erdem, A., Babic, D., & Silvano, C. (2019) · IEEE International Conference on Electronics, Circuits and Systems (ICECS) · doi:10.1109/icecs46596.2019.8964870 · Cited by 5
“We considered implementing the work [3] as well, however it was not possible to fit it on a low-density FPGA as the one targeted in this paper.”
BackgroundMethodInfluentialFPGA based convolution and memory architecture for Convolutional Neural Network
Shahan, K. A., & Rani, J. S. (2020) · International Conference on VLSI Design · doi:10.1109/vlsid49098.2020.00049 · Cited by 5
Edge Computing for Artificial Intelligence
Wang, X., Han, Y., Leung, V. C. M., Niyato, D. T., Yan, X., & Chen, X. (2020) · Springer eBooks · doi:10.1007/978-981-15-6186-3_7 · Cited by 5
ConformalLayers: A non-linear sequential neural network with associative layers
Vera, E., Fernandes, L. A. F., & Vasconcelos, C. N. (2021) · SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI) · doi:10.1109/sibgrapi54419.2021.00059 · Cited by 5
Low-memory convolutional neural networks through incremental depth-first processingPreprint
Binas, J., & Bengio, Y. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1804.10727 · Cited by 5
“Alwani et al. (2016) propose an updating scheme, where a local variable in the network is computed based on all inputs affecting this variable.”
BackgroundDNN Feature Map Compression using Learned Representation over GF(2)Preprint
Gudovskiy, D., Hodgkinson, A., & Rigazio, L. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1808.05285 · Cited by 5
“One obvious solution is to keep only a subset of intermediate feature maps at the expense of recomputing convolutions [1].”
BackgroundEnergy Predictive Models for Convolutional Neural Networks on Mobile PlatformsPreprint
Rodrigues, C. F., Riley, G. D., & Luján, M. (2020) · arXiv (Cornell University) · doi:10.13140/rg.2.2.15224.80644 · Cited by 5
“Finally, the predictive models in our work are built at the layerlevel and any optimizations to accelerate a layer, such as fusedlayer implementations [3], is typically done below this level of abstraction, and is thus automatically covered.”
MethodOptimizer Fusion: Efficient Training with Better Locality and ParallelismPreprint
Jiang, Z., Gu, J., Liu, M., Zhu, K., & Pan, D. Z. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2104.00237 · Cited by 5
Yang, S., Fang, C., Lin, H., Chen, X., Liu, J., & Chen, Z. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2506.17642 · Cited by 5
“The root cause of this bug is that PyTorch Inductor attempts to fuse torch.nn.ReflectionPad3d , view , and torch.chunk in the computational graph (this optimization is also called operator fusion [5]).”
LoopTree: Enabling Systematic and Flexible Exploration of Fused-layer Dataflow Accelerators
Gilbert, M. (2023) · DSpace@MIT (Massachusetts Institute of Technology) · Cited by 5
A CNN-Based Super-Resolution Processor With Short-Term Caching for Real-Time UHD Upscaling
Ahn, H. K., Lee, S., & Jung, S. (2023) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3346440 · Cited by 4
“The layer fusion method is proposed to resolve the memory problem caused by large FM in CNN-based algorithms [4].”
Methodmsf-CNN: Patch-based Multi-Stage Fusion with Convolutional Neural Networks for TinyML
Huang, Z., & Baccelli, E. (2025) · neural information processing systems · doi:10.52202/085713-2415 · Cited by 4
Using branch predictors to predict brain activity in brain-machine implants
Bhattacharjee, A. (2017) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1145/3123939.3123943 · Cited by 4
“In the context of studies from the architecture domain, our work relates to hardware neural network accelerators [4, 12, 13, 37], but is closer to studies linking neural biologywith architecture.”
BackgroundIn Search of the Performance- and Energy-Efficient CNN Accelerators
Sedukhin, S. G., Tomioka, Y., & Yamamoto, K. (2021) · IEICE Transactions on Electronics · doi:10.1587/transele.2021lhp0003 · Cited by 4
Mitigating Memory Wall Effects in CNN Engines with On-the-Fly Weights Generation
Venieris, S. I., Fernández-Marqués, J., & Lane, N. D. (2023) · ACM Transactions on Design Automation of Electronic Systems · doi:10.1145/3611673 · Cited by 5
“As such, unzipFPGA is orthogonal and complementary to quantisation [34], activations’ encoding [20, 68], fusion [9] and zero-skipping PEs [5, 6, 32, 56, 106].”
BackgroundLin, J.-Y., & Lin, S.-Y. (2023) · IEEE Transactions on Computers · doi:10.1109/tc.2023.3269696 · Cited by 4
“In the state-of-the-art CNNs, the accuracy of the inference is very high for the task of the computer vision [1], [2].”
BackgroundFIACCEL: Memory Efficient Frame Interpolation Accelerator for Full-HD Video
Jeong, M. W., & Rhee, C. E. (2023) · IEEE Transactions on Circuits & Systems II Express Briefs · doi:10.1109/tcsii.2023.3329966 · Cited by 4
“To reduce DRAM access in CNN-based accelerators, the fused-layer approach by Alwani et al. is a notable method [8].”
MethodDyPipe: A Holistic Approach to Accelerating Dynamic Neural Networks with Dynamic Pipelining
Zhuang, Y., Hu, X., Chen, X., & Zhi, T. (2023) · Journal of Computer Science and Technology · doi:10.1007/s11390-021-1161-y · Cited by 4
Wu, H., Chen, Y., Yuan, Y., Yue, J., Fu, X., Ren, Q., Luo, Q., Mak, P., Wang, X., & Zhang, F. (2023) · IEEE Transactions on Circuits and Systems I Regular Papers · doi:10.1109/tcsi.2023.3325850 · Cited by 4
“In summary, using the various technologies proposed in this work can solve the inherent problems in the SR task and three derived problems caused by the solutions in previous works [11], [20].”
MethodDeep neural networks accelerators with focus on tensor processors
Bolhasani, H., & Marandinejad, M. (2023) · Microprocessors and Microsystems · doi:10.1016/j.micpro.2023.105005 · Cited by 4
BrickDL: Graph-Level Optimizations for DNNs with Fine-Grained Data Blocking on GPUs
Lakshminarasimhan, M., Hall, M. W., Williams, S., & Antepara, O. (2024) · Proceedings of the International Conference on Parallel Processing · doi:10.1145/3673038.3673046 · Cited by 4
“In essence, they perturb the cache oblivious recursion algorithm, incentivizing the execution of temporally concurrent independent tiles.”
Glint, T., Pechimuthu, M., & Mekie, J. (2024) · Design, Automation, and Test in Europe · doi:10.23919/date58400.2024.10546624 · Cited by 4
“With the continuous advancement of deep learning and the need for efficient computations, frameworks like DeepFrack will play a crucial role in shaping the future of AI hardware accelerators.”
MethodSparsepipe: Sparse Inter-operator Dataflow Architecture with Cross-Iteration Reuse
Zhang, Y., Tsai, P.-A., & Tseng, H. (2024) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro61859.2024.00090 · Cited by 4
“Examples include FLAT [32] for the multi-head attention kernel in Transformers, Fused-Layer CNN Accelerators [3] for CNNs and Pipelayer [54], TANGRAM [20], ARCHON [43], and Atomic dataflow [73], for producer-consumer reuse across spatial hardware units.”
LrGAN: A Compact and Energy Efficient PIM-Based Architecture for GAN Training
Mao, H., Shu, J., Song, M., & Li, T. (2020) · IEEE Transactions on Computers · doi:10.1109/tc.2020.3011122 · Cited by 4
“Many recent works accelerate NN based on FPGAs [3], [47], [60], [72], [74] and ASICs [2], [12], [19], [24], [41], [44], [45], [58], [73].”
Background
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.