报告题目：面向不规则计算的硬件优化：稀疏矩阵加速与近存计算架构

演讲摘要：Sparse computations on matrix/tensor data are widely used in deep learning and graph applications, but efficiently supporting sparse computations on existing systems is challenging due to irregular data distribution and poor access locality. We argue that, since the irregular data distribution is difficult to predict beforehand, it would be necessary to make the hardware architecture aware of and dynamically adaptive to the specific workload at runtime. We present two example designs in this talk. The first one is a sparse matrix accelerator that uses adaptive dataflow based on data sparsity patterns. The second one is a near-data processing system that co-optimizes remote access and load imbalance issues. These two designs are both published in ASPLOS 2023.

讲者简介：清华大学交叉信息研究院助理教授，博士生导师。于美国斯坦福大学电子工程系获博士学位。研究方向为计算机体系结构与系统，主要关注针对人工智能和大数据分析等数据密集型应用的新型存储架构、专用计算系统、硬件系统安全等方向。已发表多篇国际顶级学术会议（ISCA、ASPLOS、HPCA、OSDI、PACT等）论文，曾获得IEEE Micro 2016年度计算机系统结构最佳论文奖（Top Picks）、福布斯中国2019年30 Under 30等荣誉。他也是MICRO、ASPLOS、ISCA等多个会议的TPC和ERC委员会成员。