{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,7,9]],"date-time":"2026-07-09T22:32:47Z","timestamp":1783636367952,"version":"3.55.0"},"reference-count":42,"publisher":"Association for Computing Machinery (ACM)","issue":"3","license":[{"start":{"date-parts":[[2024,5,11]],"date-time":"2024-05-11T00:00:00Z","timestamp":1715385600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/www.acm.org\/publications\/policies\/copyright_policy#Background"}],"funder":[{"name":"National Science Foundation","award":["CNS-1738783"],"award-info":[{"award-number":["CNS-1738783"]}]}],"content-domain":{"domain":["dl.acm.org"],"crossmark-restriction":true},"short-container-title":["ACM Trans. Embed. Comput. Syst."],"published-print":{"date-parts":[[2024,5,31]]},"abstract":"<jats:p>Modern deep-learning models tend to include billions of parameters, reducing real-time performance. Embedded systems are compute-constrained while frequently used to deploy these models for real-time systems given size, weight, and power requirements. Tools like parameter-scaling methods help to shrink models to ease deployment. This research compares two scaling methods for convolutional neural networks, uniform scaling and NeuralScale, and analyzes their impact on inference latency, memory utilization, and power. Uniform scaling scales the number of filters evenly across a network. NeuralScale adaptively scales the model to theoretically achieve the highest accuracy for a target parameter count. In this study, VGG-11, MobileNetV2, and ResNet-50 models were scaled to four ratios: 0.25\u00d7, 0.50\u00d7, 0.75\u00d7, 1.00\u00d7. These models were benchmarked on an ARM Cortex-A72 CPU, an NVIDIA Jetson AGX Xavier GPU, and a Xilinx ZCU104 FPGA. Additionally, quantization was applied to meet real-time objectives. The CIFAR-10 and tinyImageNet datasets were studied. On CIFAR-10, NeuralScale creates more computationally intensive models than uniform scaling for the same parameter count, with relative speeds of 41% on the CPU, 72% on the GPU, and 96% on the FPGA. The additional computational complexity is a tradeoff for accuracy improvements in VGG-11 and MobileNetV2 NeuralScale models but reduced ResNet-50 NeuralScale accuracy. Furthermore, quantization alone achieves similar or better performance on the CPU and GPU devices when compared with models scaled to 0.50\u00d7, despite slight reductions in accuracy. On the GPU, quantization reduces latency by 2.7\u00d7 and memory consumption by 4.3\u00d7. Uniform-scaling models are 1.8\u00d7 faster and use 2.8\u00d7 less memory. NeuralScale reduces latency by 1.3\u00d7 and dropped memory by 1.1\u00d7. We find quantization to be a practical first tool for improved performance. Uniform scaling can easily be applied for additional improvements. NeuralScale may improve accuracy but tends to negatively impact performance, so more care must be taken with it.<\/jats:p>","DOI":"10.1145\/3654799","type":"journal-article","created":{"date-parts":[[2024,3,30]],"date-time":"2024-03-30T09:24:03Z","timestamp":1711790643000},"page":"1-35","update-policy":"https:\/\/doi.org\/10.1145\/crossmark-policy","source":"Crossref","is-referenced-by-count":2,"title":["Characterizing Parameter Scaling with Quantization for Deployment of CNNs on Real-Time Systems"],"prefix":"10.1145","volume":"23","author":[{"ORCID":"https:\/\/orcid.org\/0000-0003-1173-0378","authenticated-orcid":false,"given":"Calvin B.","family":"Gealy","sequence":"first","affiliation":[{"name":"University of Pittsburgh, Pittsburgh, United States"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-9665-2879","authenticated-orcid":false,"given":"Alan D.","family":"George","sequence":"additional","affiliation":[{"name":"University of Pittsburgh, Pittsburgh, United States"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"320","published-online":{"date-parts":[[2024,5,11]]},"reference":[{"key":"e_1_3_2_2_2","unstructured":"ONNX Developers. 2019. ONNX: Open Neural Network Exchange. Retrieved from https:\/\/github.com\/onnx\/onnx"},{"key":"e_1_3_2_3_2","first-page":"1","volume-title":"Proceedings of the International Conference on Learning Representations","author":"Cai Han","year":"2020","unstructured":"Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. 2020. Once-for-all: Train One network and specialize it for efficient deployment. In Proceedings of the International Conference on Learning Representations. International Conference on Learning Representations, ICLR, Addis Ababa, Ethiopia, 1\u201315. Retrieved from https:\/\/github.com\/mit-han-lab\/once-for-all"},{"key":"e_1_3_2_4_2","first-page":"1","volume-title":"Proceedings of the 7th International Conference on Learning Representations, ICLR 2019","author":"Cai Han","year":"2019","unstructured":"Han Cai, Ligeng Zhu, and Song Han. 2019. Proxylessnas: Direct neural architecture search on target task and hardware. In Proceedings of the 7th International Conference on Learning Representations, ICLR 2019. International Conference on Learning Representations, ICLR, New Orleans, Louisiana, USA, 1\u201313."},{"key":"e_1_3_2_5_2","unstructured":"Tianqi Chen Ian Goodfellow and Jonathon Shlens. 2016. Net2Net: Accelerating learning via knowledge transfer. In Proceedings of the 4th International Conference on Learning Representations (ICLR\u201916). International Conference on Learning Representations ICLR San Juan Puerto Rico 1\u201312."},{"key":"e_1_3_2_6_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2009.5206848"},{"key":"e_1_3_2_7_2","first-page":"2148","volume-title":"Advances in Neural Information Processing Systems","author":"Denil Misha","year":"2013","unstructured":"Misha Denil, Babak Shakibi, Laurent Dinh, MarcAurelio Ranzato, and Nando de Freitas. 2013. Predicting parameters in deep learning. In Advances in Neural Information Processing Systems. C.J. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K.Q. Weinberger (Eds.), Vol. 26. Curran Associates, Inc., Lake Tahoe, 2148\u20132156. Retrieved from https:\/\/proceedings.neurips.cc\/paper_files\/paper\/2013\/file\/7fec306d1e665bc9c748b5d2b99a6e97-Paper.pdf"},{"key":"e_1_3_2_8_2","doi-asserted-by":"publisher","DOI":"10.1109\/ICCV48922.2021.00447"},{"key":"e_1_3_2_9_2","doi-asserted-by":"publisher","DOI":"10.1109\/TPAMI.2020.3044416"},{"key":"e_1_3_2_10_2","first-page":"1","volume-title":"Proceedings of the AIAA\/USU Conference on Small Satellites","author":"Gealy Calvin B.","year":"2022","unstructured":"Calvin B. Gealy and Alan D. George. 2022. Evaluation of parameter-scaling for efficient deep learning on small satellites. In Proceedings of the AIAA\/USU Conference on Small Satellites. USU, Logan, Utah, 1\u201314. Retrieved from https:\/\/digitalcommons.usu.edu\/smallsat\/2022\/all2022\/348\/"},{"key":"e_1_3_2_11_2","doi-asserted-by":"publisher","DOI":"10.1201\/9781003162810-13"},{"key":"e_1_3_2_12_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2016.90"},{"key":"e_1_3_2_13_2","unstructured":"Andrew G. Howard Menglong Zhu Bo Chen Dmitry Kalenichenko Weijun Wang Tobias Weyand Marco Andreetto and Hartwig Adam. 2017. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv:1704.04861. Retrieved from https:\/\/arxiv.org\/abs\/1704.04861"},{"key":"e_1_3_2_14_2","doi-asserted-by":"publisher","DOI":"10.1145\/3373087.3375887"},{"key":"e_1_3_2_15_2","doi-asserted-by":"publisher","DOI":"10.1007\/978-3-030-01267-0_35"},{"key":"e_1_3_2_16_2","unstructured":"Alex Krizhevsky Vinod Nair and Geoffrey Hinton. 2009. CIFAR-10 and CIFAR-100 datasets. Retrieved from https:\/\/www.cs.toronto.edu\/kriz\/cifar.html"},{"key":"e_1_3_2_17_2","doi-asserted-by":"publisher","DOI":"10.1109\/HPEC43674.2020.9286182"},{"key":"e_1_3_2_18_2","unstructured":"Eugene Lee and Chen-Yi Lee. 2020. GitHub\u2014NeuralScale: Efficient Scaling of Neurons for Resource-Constrained Deep Neural Networks. Retrieved from https:\/\/github.com\/eugenelet\/NeuralScale"},{"key":"e_1_3_2_19_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR42600.2020.00155"},{"key":"e_1_3_2_20_2","unstructured":"Fei-Fei Li Andrej Karpathy and Justin Johnson. 2017. tinyImageNet from Stanford University CS231n. Retrieved from http:\/\/cs231n.stanford.edu\/tiny-imagenet-200.zip"},{"key":"e_1_3_2_21_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR46437.2021.00865"},{"key":"e_1_3_2_22_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2019.01152"},{"key":"e_1_3_2_23_2","unstructured":"NVIDIA. 2021. Developer Guide\u2014NVIDIA Deep Learning TensorRT Documentation v8.0.3. Retrieved from https:\/\/docs.nvidia.com\/deeplearning\/tensorrt\/archives\/tensorrt-803\/developer-guide\/index.html"},{"key":"e_1_3_2_24_2","unstructured":"NVIDIA. 2022. Developer Guide :: NVIDIA Deep Learning TensorRT Documentation v8.5.2. Retrieved from https:\/\/docs.nvidia.com\/deeplearning\/tensorrt\/archives\/tensorrt-852\/developer-guide\/index.html#intro-quantization"},{"key":"e_1_3_2_25_2","unstructured":"NVIDIA. 2023. GeForce GTX 1080 Ti Graphics Cards. Retrieved from https:\/\/www.nvidia.com\/en-in\/geforce\/products\/10series\/geforce-gtx-1080-ti\/"},{"key":"e_1_3_2_26_2","unstructured":"NVIDIA. 2023. Jetson Modules. Retrieved from https:\/\/developer.nvidia.com\/embedded\/jetson-modules"},{"key":"e_1_3_2_27_2","unstructured":"NVIDIA. 2023. ONNX-TensorRT: TensorRT Backend For ONNX. Retrieved from https:\/\/github.com\/onnx\/onnx-tensorrt"},{"key":"e_1_3_2_28_2","unstructured":"ONNX Runtime developers. 2021. ONNX Runtime. Retrieved from https:\/\/onnxruntime.ai\/"},{"key":"e_1_3_2_29_2","unstructured":"ONNX Runtime developers. 2023. Quantize ONNX Models. Retrieved from https:\/\/onnxruntime.ai\/docs\/performance\/quantization.html"},{"key":"e_1_3_2_30_2","unstructured":"Fabian Pedregosa and Philippe Gervais. 2022. Memory_profiler: Monitor Memory Usage of Python Code. Retrieved from https:\/\/github.com\/pythonprofilers\/memory_profiler"},{"key":"e_1_3_2_31_2","unstructured":"Facundo Quiroga. 2023. Facundoq\/Tinyimagenet. Retrieved from https:\/\/github.com\/facundoq\/tinyimagenet"},{"key":"e_1_3_2_32_2","unstructured":"Raspberry Pi Foundation. 2023. Raspberry Pi 4 Model B Specifications\u2014Raspberry Pi. Retrieved from https:\/\/www.raspberrypi.com\/products\/raspberry-pi-4-model-b\/specifications\/"},{"key":"e_1_3_2_33_2","doi-asserted-by":"publisher","DOI":"10.1109\/CVPR.2018.00474"},{"key":"e_1_3_2_34_2","first-page":"1","volume-title":"Proceedings of the 3rd International Conference on Learning Representations, ICLR 2015","author":"Simonyan Karen","year":"2015","unstructured":"Karen Simonyan and Andrew Zisserman. 2015. Very deep convolutional networks for large-scale image recognition. In Proceedings of the 3rd International Conference on Learning Representations, ICLR 2015. International Conference on Learning Representations, ICLR, San Diego, California, USA, 1\u201314."},{"key":"e_1_3_2_35_2","unstructured":"TechPowerUp. 2023. NVIDIA GeForce GTX 1080 Ti Specs. Retrieved from https:\/\/www.techpowerup.com\/gpu-specs\/geforce-gtx-1080-ti.c2877"},{"key":"e_1_3_2_36_2","unstructured":"TechPowerUp. 2023. NVIDIA Jetson AGX Xavier GPU Specs. Retrieved from https:\/\/www.techpowerup.com\/gpu-specs\/jetson-agx-xavier-gpu.c3232"},{"key":"e_1_3_2_37_2","unstructured":"Xilinx. 2022. Vitis-AI-Tutorials\/Design_Tutorials\/09-Mnist_p yt at 1.4 \\(\\cdot\\) Xilinx\/Vitis-AI-Tutorials. Retrieved from https:\/\/github.com\/Xilinx\/Vitis-AI-Tutorials"},{"key":"e_1_3_2_38_2","unstructured":"Xilinx. 2023. IP Facts \u2022 DPUCZDX8G for Zynq UltraScale+ MPSoCs Product Guide (PG338) \u2022 Reader \u2022 AMD Adaptive Computing Documentation Portal. Retrieved from https:\/\/docs.xilinx.com\/r\/en-US\/pg338-dpu\/IP-Facts"},{"key":"e_1_3_2_39_2","unstructured":"Xilinx. 2023. Versal AI Core Series. Retrieved from https:\/\/www.xilinx.com\/products\/silicon-devices\/acap\/versal-ai-core.html"},{"key":"e_1_3_2_40_2","unstructured":"Xilinx. 2023. Xilinx\/Vitis-AI at v3.0. Retrieved from https:\/\/github.com\/Xilinx\/Vitis-AI\/tree\/v3.0"},{"key":"e_1_3_2_41_2","unstructured":"Xilinx. 2023. Zynq UltraScale+ MPSoC ZCU104 Evaluation Kit. Retrieved from https:\/\/www.xilinx.com\/products\/boards-and-kits\/zcu104.html"},{"key":"e_1_3_2_42_2","doi-asserted-by":"publisher","DOI":"10.1007\/978-3-319-10590-1_53"},{"key":"e_1_3_2_43_2","unstructured":"Ligeng Zhu. 2022. PyTorch-OpCounter. Retrieved from https:\/\/github.com\/Lyken17\/pytorch-OpCounter"}],"container-title":["ACM Transactions on Embedded Computing Systems"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/dl.acm.org\/doi\/10.1145\/3654799","content-type":"unspecified","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/dl.acm.org\/doi\/pdf\/10.1145\/3654799","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2025,6,19]],"date-time":"2025-06-19T00:06:09Z","timestamp":1750291569000},"score":1,"resource":{"primary":{"URL":"https:\/\/dl.acm.org\/doi\/10.1145\/3654799"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2024,5,11]]},"references-count":42,"journal-issue":{"issue":"3","published-print":{"date-parts":[[2024,5,31]]}},"alternative-id":["10.1145\/3654799"],"URL":"https:\/\/doi.org\/10.1145\/3654799","relation":{},"ISSN":["1539-9087","1558-3465"],"issn-type":[{"value":"1539-9087","type":"print"},{"value":"1558-3465","type":"electronic"}],"subject":[],"published":{"date-parts":[[2024,5,11]]},"assertion":[{"value":"2023-10-10","order":0,"name":"received","label":"Received","group":{"name":"publication_history","label":"Publication History"}},{"value":"2024-03-21","order":1,"name":"accepted","label":"Accepted","group":{"name":"publication_history","label":"Publication History"}},{"value":"2024-05-11","order":2,"name":"published","label":"Published","group":{"name":"publication_history","label":"Publication History"}}]}}