Abstract:
As Moore's law reaches the end of life, there is an urgency to find alternative computing method to meet the soaring demands of modern computing, e.g., that from machine learning applications. In this talk, I will elaborate our efforts on using "time" as a new dimension for accelerating computing in modern CMOS integrated circuits (ICs). I will first discuss our developments on mixed-signal time domain computing where we use "time" as information carrier to solve modern machine learning problems such as time series analysis, image recognition, and artificial neural network. Second, I will show how to leverage the instruction level timing information in GPU processors to achieve energy saving on modern computing tasks including machine learning jobs.