{"links":{},"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00237576","sets":["1164:2240:11467:11704"]},"path":["11704"],"owner":"44499","recid":"237576","title":["Enhancing Sparse DNN Inference on GPUs: Adaptive Tile Pruning and Split-Tiled Sparse Matrix Multiplication"],"pubdate":{"attribute_name":"公開日","attribute_value":"2024-08-01"},"_buckets":{"deposit":"c3decbb7-3b0c-48f7-8393-65c522a3f688"},"_deposit":{"id":"237576","pid":{"type":"depid","value":"237576","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"Enhancing Sparse DNN Inference on GPUs: Adaptive Tile Pruning and Split-Tiled Sparse Matrix Multiplication","author_link":["650573","650572","650571","650574"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"Enhancing Sparse DNN Inference on GPUs: Adaptive Tile Pruning and Split-Tiled Sparse Matrix Multiplication"},{"subitem_title":"Enhancing Sparse DNN Inference on GPUs: Adaptive Tile Pruning and Split-Tiled Sparse Matrix Multiplication","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"深層学習","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2024-08-01","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"Graduate School of Information Science and Technology, Osaka University"},{"subitem_text_value":"Graduate School of Information Science and Technology, Osaka University"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Graduate School of Information Science and Technology, Osaka University","subitem_text_language":"en"},{"subitem_text_value":"Graduate School of Information Science and Technology, Osaka University","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"eng"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/237576/files/IPSJ-HPC24195015.pdf","label":"IPSJ-HPC24195015.pdf"},"date":[{"dateType":"Available","dateValue":"2026-08-01"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-HPC24195015.pdf","filesize":[{"value":"2.1 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"14"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"12e088c6-b88b-48d1-9262-a71f4a08c808","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2024 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Yanchen, Li"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Fumihiko, Ino"}],"nameIdentifiers":[{}]}]},"item_4_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Yanchen, Li","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Fumihiko, Ino","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10463942","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8841","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"Deep neural network (DNN) pruning is a popular method for accelerating computations in DNNs by removing unimportant parameters. Among pruning methods, tile-wise pruning (TWP) with sparse matrix multiplication achieves significant acceleration with minimal pruning loss. However, sparse matrix multiplication based on TWP suffers from load imbalance when important weight elements in the matrices of the DNN are unevenly distributed. To address this issue, we propose Adaptive Tile Pruning (ATP) and Split-Tiled Sparse Matrix Multiplication (STSpMM). ATP constructs sparse matrices with flexibly balanced workloads while preserving DNN model accuracy. Meanwhile, STSpMM efficiently handles ATP-generated sparse matrices on GPUs by splitting and redistributing large workloads. We evaluated our approach on pruned ResNet-34 model using ImageNet, and BERT-Small on QNLI tasks. Results demonstrate that ATP-pruned models processed via STSpMM achieve greater acceleration than previous methods while maintaining accuracy.","subitem_description_type":"Other"}]},"item_4_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"Deep neural network (DNN) pruning is a popular method for accelerating computations in DNNs by removing unimportant parameters. Among pruning methods, tile-wise pruning (TWP) with sparse matrix multiplication achieves significant acceleration with minimal pruning loss. However, sparse matrix multiplication based on TWP suffers from load imbalance when important weight elements in the matrices of the DNN are unevenly distributed. To address this issue, we propose Adaptive Tile Pruning (ATP) and Split-Tiled Sparse Matrix Multiplication (STSpMM). ATP constructs sparse matrices with flexibly balanced workloads while preserving DNN model accuracy. Meanwhile, STSpMM efficiently handles ATP-generated sparse matrices on GPUs by splitting and redistributing large workloads. We evaluated our approach on pruned ResNet-34 model using ImageNet, and BERT-Small on QNLI tasks. Results demonstrate that ATP-pruned models processed via STSpMM achieve greater acceleration than previous methods while maintaining accuracy.","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"8","bibliographic_titles":[{"bibliographic_title":"研究報告ハイパフォーマンスコンピューティング（HPC）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2024-08-01","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"15","bibliographicVolumeNumber":"2024-HPC-195"}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"updated":"2025-01-19T08:50:21.026483+00:00","created":"2025-01-19T01:40:16.875852+00:00","id":237576}