Layer-wise power/performance modelling for single-board CNN inference

• Item type: SIG Technical Reports(1)
• 公開日: 2022-03-03
• タイトル: Layer-wise power/performance modelling for single-board CNN inference
• タイトル / 言語: en
• タイトル: Layer-wise power/performance modelling for single-board CNN inference
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: ニューラルネットワーク
• 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: Kyushu Uniersity
• 著者所属: Kyushu Uniersity
• 著者所属: Kyushu Uniersity
• 著者所属: Kyushu Uniersity
• 著者所属: Kyushu Uniersity
• 著者所属(英): en
• 著者所属(英): Kyushu Uniersity
• 著者所属(英): en
• 著者所属(英): Kyushu Uniersity
• 著者所属(英): en
• 著者所属(英): Kyushu Uniersity
• 著者所属(英): en
• 著者所属(英): Kyushu Uniersity
• 著者所属(英): en
• 著者所属(英): Kyushu Uniersity
• 著者名: Yi, Ng Kuan; Aalaa, M.A. Babai; Satoshi, Kawakami; Teruo, Tanimoto; Koji, Inoue
• 著者名(英): Yi, Ng Kuan; Aalaa, M.A. Babai; Satoshi, Kawakami; Teruo, Tanimoto; Koji, Inoue
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Intermittent executions and energy harvesting technologies are promising candidates to enable renewable energy on small-scale computer systems like single-board computers, making sustainable computing possible. In this work, we implemented an energy consumption prediction framework for each layer of CNN executing on single-board computers based on NeuralPower as the first step towards enabling energy-efficient intermittent execution of CNN inference on single-board computers. We found that layer hyperparameters cannot explain all the variations in execution time and power consumption when the layer is executed. Model's prediction can be improved with the knowledge of performance counter values, but these values are not available before a layer is executed. Furthermore, our analysis revealed that implementation optimization like sparse matrix multiplication might cause a layer's execution time and power to change with its input values.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Intermittent executions and energy harvesting technologies are promising candidates to enable renewable energy on small-scale computer systems like single-board computers, making sustainable computing possible. In this work, we implemented an energy consumption prediction framework for each layer of CNN executing on single-board computers based on NeuralPower as the first step towards enabling energy-efficient intermittent execution of CNN inference on single-board computers. We found that layer hyperparameters cannot explain all the variations in execution time and power consumption when the layer is executed. Model's prediction can be improved with the knowledge of performance counter values, but these values are not available before a layer is executed. Furthermore, our analysis revealed that implementation optimization like sparse matrix multiplication might cause a layer's execution time and power to change with its input values.
• 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: AN10096105
• 書誌情報: 研究報告システム・アーキテクチャ（ARC）; 巻 2022-ARC-248, 号 13, p. 1-11, 発行日 2022-03-03
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8574
• 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