The End of Moore's Law and the Future of Computing Systems, Probably

摘要

Many claim that the laws of physics dictating the exponentially improving benefits of Moore's Law will end in the next 10 to 20 years. The argument for the end of Moore's Law is based, in part, on an analysis that switching devices cannot function deterministically as feature sizes get reduced to molecular levels. Moore'Bs Law could, however, continue provided systems with probabilistic switches could process information usefully. My research suggests that this is indeed possible in contexts where the "quality" of the results of the computation is perceptually determined by our senses audio and video information being significant examples. To demonstrate this principle, I will show how CMOS-based devices, circuits and computing architectures whose correctness is characterized probabilistically can be used effectively. I will show that significant (a multiplicative factor of 2 or more) energy and performance gains can be achieved, while trading a perceptually tolerable level of error?through probabilistic adders and multipliers, applied in the context of finite impulse response filters and FFT engines processing video and audio data in digital signal processing. Quantifying the human tolerance for error, we expect, will be ultimately based on neurobiological models. Conceptually, our thesis recommending tolerating error in the switching devices in return for savings in cost, is analogous to Simon ' s notion if satisficing?we will conclude the talk by dwelling on this analog and its implications to probabilistic design of ultra large-scale integrated (ULSI) circuits in the future.