{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/escort-efficient-sparse-convolutional-neural","title":"Escoin: Efficient Sparse Convolutional Neural Network Inference on GPUs","arxiv_id":"1802.10280","date":"2018-02-28","proceeding":null,"authors":["Xuhao Chen"],"abstract":"Deep neural networks have achieved remarkable accuracy in many artificial\nintelligence applications, e.g. computer vision, at the cost of a large number\nof parameters and high computational complexity. Weight pruning can compress\nDNN models by removing redundant parameters in the networks, but it brings\nsparsity in the weight matrix, and therefore makes the computation inefficient\non GPUs. Although pruning can remove more than 80% of the weights, it actually\nhurts inference performance (speed) when running models on GPUs.\n  Two major problems cause this unsatisfactory performance on GPUs. First,\nlowering convolution onto matrix multiplication reduces data reuse\nopportunities and wastes memory bandwidth. Second, the sparsity brought by\npruning makes the computation irregular, which leads to inefficiency when\nrunning on massively parallel GPUs. To overcome these two limitations, we\npropose Escort, an efficient sparse convolutional neural networks on GPUs.\nInstead of using the lowering method, we choose to compute the sparse\nconvolutions directly. We then orchestrate the parallelism and locality for the\ndirect sparse convolution kernel, and apply customized optimization techniques\nto further improve performance. Evaluation on NVIDIA GPUs show that Escort can\nimprove sparse convolution speed by 2.63x and 3.07x, and inference speed by\n1.43x and 1.69x, compared to CUBLAS and CUSPARSE respectively.","url_abs":"http://arxiv.org/abs/1802.10280v2","url_pdf":"http://arxiv.org/pdf/1802.10280v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"escort-efficient-sparse-convolutional-neural","repo_url":"https://github.com/chenxuhao/caffe-escoin","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok","spdx":"NOASSERTION"}}],"tasks":[],"methods":[{"method_slug":"convolution","method_name":"Convolution"},{"method_slug":"pruning","method_name":"Pruning"},{"method_slug":"speed","method_name":"SPEED"}],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1802.10280","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}