| 赤部 知也 | D, 中間発表 | コンピューティング・アーキテクチャ | 中島 康彦, 林 優一, 張 任遠, KAN Yirong, PHAM HOAI LUAN |
|
title: Evaluation of Multilevel Pipeline CGRA with Double Buffer and Dual Address Synchronizer
abstract: Based on IMAX2, which has a ring structure, column multi-threading, DMA-informed load/store, and 5-input conditional operations to ease CGRA's positional constraints, we propose the following (1) The ability to use the cache as a double buffer to handle data flow splitting in FFT and merge sorting. (2) An address tuning mechanism in the basic PE to support sparse matrix product (SpGEMM) and merge sorting. Based on the above ideas, this paper proposes IMAX3: a multi-level pipelined CGRA. The basic performance of individual CGRAs connected to the multi-port main memory was evaluated using three application programs (SpGEMM, FFT, and merge sort), and the results show better performance per bandwidth and per power than GPUs. language of the presentation: Japanese 発表題目: ダブルバッファとアドレス同調機構を備えるマルチレベル・パイプラインCGRAの評価 発表概要: CGRAの位置制約を緩和するリング構造，列マルチスレッド，DMA情報付きロード/ストア，5入力条件付き演算を備えるIMAX2をベースとして以下を提案する． (1) FFTとマージソートにおけるデータフローの分割を処理するために，キャッシュをダブルバッファとして使用する機能を搭載． (2) 疎行列積(SpGEMM)とマージソートをサポートするために，基本PEにアドレス同調機構を搭載． 本論文では，上記の考え方に基づき，IMAX3：マルチレベルパイプライン型CGRAを提案する． マルチポートメインメモリに接続された個々のCGRAの基本性能を，3つのアプリケーションプログラム（SpGEMM，FFT，マージソート）を用いて評価した結果， GPUより優れたバンド幅あたりの性能，電力あたりの性能を示した． | |||
| KIM DOHYUN | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, 林 優一, 張 任遠, KAN Yirong, PHAM HOAI LUAN |
|
title: Graph Convolution Network Acceleration by using IMAX2
abstract: Graph Convolution Network, in shortly, GCN is a neural network architecture for analyze some graph-structured data. GCN is a versatile model becuase GCN is not only used in natural science divisions such as analyze molcular or proteins, but also used in analyze relationship of sales data or papers data. However, GCN is also required high computing power like other neural network models so it is not good power efficiency what calculating the model on traditional computation infrastructres. For solving this problem, I aim to improve power efficiency than traditional computation base such as CPU or GPU by using IMAX2 that is Coarse-Grain Reconfigurable Architecture(CGRA). In this presentation, I will describe about research backgrounds and the purpose, related researchs, previous works, results obtained from this research, future works. language of the presentation: Japanese 発表題目: IMAX2を用いたGraph Convolutional Networkアクセラレーション 発表概要: Graph Convolution Network、略してGCNはグラフ構造のデータの分析に用いられるニューラルネットワークアーキテクチャである。GCNは分子やたんぱく質の分析等の自然科学分野に限らず、セールスデータや論文データの関係分析等にも用いられる汎用性の高いモデルである。しかし、GCNは他のニューラルネットワークのように多大な演算能力を要し、既存の計算基盤では電力効率が悪い。この問題を解決するため、粗粒度再構成可能アーキテクチャの一種であるIMAX2を用い、CPUやGPUなどの既存の計算基盤より電力効率を改善することを目標とする。本発表では、研究背景や目的、関係研究や先行研究、得られた成果、そして今後の計画について述べる。 | |||
| 今村 廉 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, 林 優一, 張 任遠, KAN Yirong, PHAM HOAI LUAN |
| title: Efficient 3D convolution using IMAX2 abstract: 3D convolutional neural networks (CNN) are widely used in medical image processing, video processing and analysis, 3D object recognition, etc. Unlike conventional image processing, 3D convolution often handles large-scale 3D data, high-resolution images, and video data. Therefore, it is important to consider not only the speed but also the energy efficiency of the computational resources used. IMAX2, a non-Von Neumann type computer developed in our laboratory, is a general-purpose computation platform with high power efficiency. In this presentation, we describe the evaluation results of 3D convolution performance by IMAX2. | |||
| 阿部 虹稀 | M, 2回目発表 | 情報セキュリティ工学 | 林 優一, 岡田 実, 藤川 和利, 藤本 大介, KIM Youngwoo（客員助教） |
|
title: Study on EM information leakage evaluation of encoded information at a distant
abstract: A threat known as electromagnetic (EM) information leakage has been reported, which involves the measurement and analysis of EM waves emitted during the operation of information devices in order to reconstruct information. Traditional EM information leakage targeted information communicated between people and devices, such as displays and keyboards. So, the information communicated was in a format easily understood by humans, and even if some of the reconstructed information was missing due to background noise, it could be interpolated from the surrounding successfully reconstructed information. However, information communicated between devices is transmitted in encoded formats that are not directly interpretable by humans. As a result, even a single bit of error in the reconstruction process could lead to error propagation throughout the data, making information reconstruction difficult. It is therefore assumed that the susceptibility to noise is higher, although the assessment of leakage from a distant location has not been previously considered. This study aims to conduct an attack assessment concerning the threat of EM information leakage for encoded information from a distant location. In this presentation, we also propose information reconstruction methods for scenarios with low leakage intensity from device-generated EM waves and demonstrate the potential for remote attacks. language of the presentation: Japanese | |||