research

Computer Architecture Design under the Impact of Process Variation:
As integrated circuit technology keeps scaling down, process variation becomes an importance factor which greatly affects the computation capability of massively parallel processors, such as GPUs and deep learning accelerators. We analyze the impact of process variation on GPU streaming processors and propose management policies to improve the energy efficiency. We also design novel deep learning accelerator architecture which is able to mitigate the impact of process variation.[ICCD'19][TODAES'19][DATE'18]

Soft Error Analysis and Fault-tolerant Computer Architecture Design:
Soft error is inevitable for integrated circuits, and it may cause wrong computation result or even system failure. We construct models to analyze and predict soft errors for GPUs and deep learning accelerators. Also, we design protection mechanism such as ECC to tolerate soft errors for GPU register files and streaming processors.[SC'21][FGCS'20][DATE'20][GLSVLSI'20]

Energy-efficient System Design for AR/VR, MAVs, etc.:
GPUs are extensively used for emerging applications in AR/VR, and these applications usually have high requirements on performance and energy consumption. Simultaneous localization and mapping (SLAM) is one of such applications, and we design a fine-grained online adaptive approximation control method for dense SLAM on embedded GPUs. Low energy consumption is essential for micro air vehicles (MAVs), and we propose an approximate execution scheme for autonomous flight in MAVs. We also design energy-efficient L2 cache by leveraging the similarity of instruction level data locality.[TODAES'21][ASAP'21][TODAES'20][DATE'19]

User Centric Computing:
Satisfying individual user is critical for user-interactive computing systems, such as mobile computers. For each individual user, we recommend the best device configuration, improve the energy efficiency of mobile devices by exploring user behavior, and design a customized hybrid power source.[TECS'20][JSA'19][PMC'19]