成果報告書詳細
管理番号20160000000254
タイトル*平成27年度中間年報 エネルギー・環境新技術先導プログラム ビッグデータ処理を加速・利活用する脳型推論システムの研究開発ー新原理デバイス・回路による超高速・低消費電力ハードウェア技術の開発とそのシステム化ー
公開日2016/4/29
報告書年度2015 - 2015
委託先名国立研究開発法人産業技術総合研究所 学校法人早稲田大学 パナソニックセミコンダクターソリューションズ株式会社 国立大学法人北海道大学
プロジェクト番号P14004
部署名イノベーション推進部
和文要約
英文要約Title:Advanced Research Program for Energy and Environmental Technologies / Artificial Neural Network System for Inference to accelerate and utilize the big data processing: Development and societal implementation of emerging devices and architectures for ultra-high-speed, low power consumption computing hardware technologies (FY2015-FY2016) FY2015 Annual Report

Material selection of a switching layer of an atomic switch is a key to achieve widening a dynamic range and decreasing power consumption. Requirements for the materials were investigated in terms of electronic conductivity and ionic conductivity, with some preliminary experiments. The specification of the electronic characterization evaluation equipment has been completed. Regarding the development of an integrated circuit of analog resistance switching devices, the leaning operation sequence has been considered assuming 0.18um devices. We have considered the neural network circuit system and confirmed the operation by using SPICE model. Towards building hardware-oriented deep learning and its system architectures for big-data processing, we built: i) a SPICE model of ReRAM (Resistance Random Access Memory) devices for integrated numerical simulations with CMOS circuits and ii) an elementary neural network architectures for deep learning. We then evaluated iii) a binary-weight (BinaryConnect) model of deep learning, and found that the strategy was significantly effective under the quantized learning algorithm. We held the R&D promotion meeting, and launched the discussion to propose the National Project.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る