Public Seminar by Advanced Materials Thrust, Function Hub , HKUST(GZ) - Robust Computing Against Unreliable Hardware

Recent breakthroughs in microelectronic scaling and artificial intelligence (AI) have brought tremendous capacity and performance advantages that continue to drive new devices and systems from high-performance computing to low-power edge devices. However, as transistor size continues to scale down to deep nanometer, microelectronic circuits are increasingly susceptible to faults/errors such as variation-induced timing errors and radiation-induced soft errors. These errors often manifest as incorrect computations which can ultimately lead to wrong results or degraded services in various applications such as AI. In this presentation, I will discuss our study on robust computing against unreliable hardware. Firstly, I will present our research on modeling microelectronic variation-induced timing errors using machine learning methods. Secondly, I will describe our research on evaluating the impact of hardware errors on AI model quality by developing PyTorch-based error injection tool. Lastly, I will discuss our efforts on improving the computing robustness against hardware errors using software and hardware solutions.