AIsim: Functional Simulator for Analog-to-Information Perceptual Systems
Hong Liu, Zheyu Liu, Fei Qiao, Mark Po-Hung Lin, Qi Wei, Huazhong Yang · 2017
Neural network accelerators have been extensively studied for artificial intelligent applications for recent years. Analogto-information systems (AIS), which could accelerate neural network computation in analog domain, are considered as one better alternative for achieving higher scalability and energy efficiency. Unlike operating in the digital domain, an AIS adopts analog computing units to construct the whole neural network for perceptual tasks. However, when designing an AIS, conventional HSPICE simulation is very time-consuming. In order to improve the design capacity, this paper presents a functional simulator, called AIsim for AIS. AIsim could efficiently simulate the noise characteristics in AIS while processing neural network algorithms. In particular, it could also guide the designers to explore the design space of analog computing units with various signal-to-noise ratio (SNR). Compared with conventional HSPICE simulator, AIsim can achieve more than 2000X simulation speedup with 2% results difference at most, for some typical benchmarks.