近似计算中基于概率图模型的软错误量化方法研究

项目名称： 近似计算中基于概率图模型的软错误量化方法研究

项目编号： No.61502298

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 自动化技术、计算机技术

项目作者： 焦佳佳

作者单位： 上海海事大学

项目金额： 20万元

中文摘要： 随着工艺尺寸的缩小，更高的软错误率和复杂的多位翻转对处理器设计带来的挑战日益严峻。同时，近似计算作为节省能耗的有效手段，也成为处理器极具前瞻性的研究方向。但是，面向近似计算的软错误量化方法尚存在空白，因此针对当前处理器近似计算中固有精度可损特性和冗余执行的在软错误分析方法中识别难题，课题拟提出基于概率图的近似量化方法，包括：1）构建基于图参数快速评估的贝叶斯网络，在适应近似计算精度可损的同时大大提高量化速度；2）制定概率图中节点参数修正策略，以满足近似计算中冗余结构的精确评估；3）设计直方图的边界分析模型，近似扩展概率图模型方法的评估结果，解决多位翻转的软错误量化难题。同时，课题拟在开源的处理器平台实现近似的概率图量化方法，并对近似计算特有的测试程序进行验证和评估。总之，课题拟研究近似概率图模型方法及其完整的实现和验证,进而为近似计算的可靠性设计提供有力的技术、方法和工具支持。

中文关键词： 近似计算；概率图模型；软错误量化；多位翻转；边界模型

英文摘要： With the scaling IC technology, increasing soft error rate and complex bit-upset patterns are challenging the reliable design. In the meanwhile, approximate computing is a compromising branch of processor due to its low power. However, Few soft error analysis methods are proposed for approximate computing. Therefore, we exploit the inexactness tolerance and selective redundant execution of approximate computing completely, and then propose a Probabilistic Graphical Models (PGM) based approximate framework. The proposed methodology includes three points: 1) constructing the corresponding Bayesian network based on fast estimation of graph parameters, which takes advantages of the acceptable inexactness of approximate computing for the high speed of estimation with guaranteed accuracy; 2) formulating the modification policy of node parameters, which satisfies the accurate estimation of redundant execution in approximate computing; 3) designing the histogram based boundary model, which extends the estimation results of proposed PGM method to the complex case of Multi-Cell Upsets. What’s more, we will finish the implementation in open source platform and verification using specified benchmarks of approximate computing. In all, such a new framework can provide the powerful technology, methodology and tool for the reliable design of approximate computing in processors.

英文关键词： Approximate Computing;Probabilistic Graphical Model (PGM);Soft error estimation;Multi-Cell Upsets (MCU);Boundary Model