{"links":{},"id":2007935,"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:02007935","sets":["1164:2240:1771568311705:1771568419073"]},"path":["1771568419073"],"owner":"80578","recid":"2007935","title":["Cerebras CS-3における密行列積アルゴリズムの性能比較"],"pubdate":{"attribute_name":"PubDate","attribute_value":"2026-03-09"},"_buckets":{"deposit":"53482db7-ad14-4f54-8ddf-4c912834a2a2"},"_deposit":{"id":"2007935","pid":{"type":"depid","value":"2007935","revision_id":0},"owners":[80578],"status":"published","created_by":80578},"item_title":"Cerebras CS-3における密行列積アルゴリズムの性能比較","author_link":[],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"Cerebras CS-3における密行列積アルゴリズムの性能比較","subitem_title_language":"ja"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"アクセラレータ計算","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2026-03-09","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"明治大学理工学部情報科学科"},{"subitem_text_value":"明治大学理工学部情報科学科"},{"subitem_text_value":"Argonne National Laboratory, Mathematics and Computer Science"},{"subitem_text_value":"明治大学理工学部情報科学科"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Argonne National Laboratory, Mathematics and Computer Science","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"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/2007935/files/IPSJ-HPC26203018.pdf","label":"IPSJ-HPC26203018.pdf"},"date":[{"dateType":"Available","dateValue":"2028-03-09"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-HPC26203018.pdf","filesize":[{"value":"3.7 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":"22623359-e4fe-47b7-9311-ece68acb9bd9","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2026 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"松崎,竜之介"}]},{"creatorNames":[{"creatorName":"村上,魁"}]},{"creatorNames":[{"creatorName":"吉井,一友"}]},{"creatorNames":[{"creatorName":"椋木,大地"}]},{"creatorNames":[{"creatorName":"宮島,敬明"}]}]},"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":"近年の計算機は演算性能が向上しているが、Byte/Flops値で表されるメモリ帯域と演算性能の比率は悪化し続けており、増大する電力性能も性能向上の阻害要因となっている。またマルチノード環境では演算に比べて非常に長いノード間通信がボトルネックになり、スケーラビリティを高めることが難しくなっている。Cerebras CS-3システム（CS-3）は、Wafer-Scale Engine 3 (WSE-3) を搭載したCerebras Systems社の最新世代のAIアクセラレータである。WSE-3は、約90万個のProcessing Element (PE) が2次元メッシュトポロジで接続され、各PEは7段パイプラインのインオーダー型プロセッサと48KBのローカルメモリで構成されている。本稿では、CS-3の大規模科学技術計算への適用可能性を理解するために、単精度浮動小数点行列積の性能評価を行なう。具体的には、SUMMAとCannonの2種類の分散並列行列積アルゴリズムについて、最大性能、強スケーリング、弱スケーリングと処理時間の内訳を示す。最大性能は、SUMMAで400.87 TFlops/s、Cannonで391.12 TFlops/sであった。サイクル数の内訳は、SUMMAで通信が53%、計算が35%、Cannonは通信が20%、計算が67%であった。弱スーリングの測定では、どちらも並列化効率が1.00であった。また、NVIDIA GH200システムでも強スケーリングの測定を行なった。","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":"2026-03-09","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"18","bibliographicVolumeNumber":"2026-HPC-203"}]},"relation_version_is_last":true,"weko_creator_id":"80578"},"created":"2026-02-25T05:52:46.507396+00:00","updated":"2026-02-25T05:52:50.426542+00:00"}