Preliminary Data-Pattern Analysis towards Energy-Efficient Adaptive In-Cache Computing for CNNs Acceleration\n

• Item type: SIG Technical Reports(1)
• 公開日: 2023-11-10
• タイトル: Preliminary Data-Pattern Analysis towards Energy-Efficient Adaptive In-Cache Computing for CNNs Acceleration\n
• タイトル / 言語: en
• タイトル: Preliminary Data-Pattern Analysis towards Energy-Efficient Adaptive In-Cache Computing for CNNs Acceleration\n
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: アーキテクチャ・CiM
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Graduate School of Information Science and Electrical Engineering, Kyushu University
• 著者所属: Faculty of Information Science and Technology, Kyushu University
• 著者所属(英): en
• 著者所属(英): Graduate School of Information Science and Electrical Engineering, Kyushu University
• 著者所属(英): en
• 著者所属(英): Faculty of Information Science and Technology, Kyushu University
• 著者名: Zhengpan, Fei; Koji, Inoue
• 著者名(英): Zhengpan, Fei; Koji, Inoue
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: In Look-Up Table (LUT) based computing, naively covering all possible results requires an exponential amount of hardware resources. While applying a decomposition technique can reduce the LUT size, it also involves considerable memory access overhead. We exploit the data pattern in Convolutional Neural Networks (CNNs) applications and propose a data-pattern- driven (DPD) optimization to drastically reduce the size of the LUT while keeping the computation efficiency. Our preliminary evaluation shows that the above scheme maintains 96.84%, 88.67%, and 76.61% of LUT coverage when reducing the LUT to 1/32, 1/64, and 1/128 of the original size, respectively.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: In Look-Up Table (LUT) based computing, naively covering all possible results requires an exponential amount of hardware resources. While applying a decomposition technique can reduce the LUT size, it also involves considerable memory access overhead. We exploit the data pattern in Convolutional Neural Networks (CNNs) applications and propose a data-pattern- driven (DPD) optimization to drastically reduce the size of the LUT while keeping the computation efficiency. Our preliminary evaluation shows that the above scheme maintains 96.84%, 88.67%, and 76.61% of LUT coverage when reducing the LUT to 1/32, 1/64, and 1/128 of the original size, respectively.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11451459
• 書誌情報: 研究報告システムとLSIの設計技術（SLDM）; 巻 2023-SLDM-204, 号 13, p. 1-6, 発行日 2023-11-10
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8639
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会