abstract bibtex

By viewing noise as a resource rather than as an impediment, we demonstrate an entirely novel approach to ultra low-energy computing. The subject of this study is the probabilistic inverter, ubiquitous to the design of digital systems, whose behavior is rendered probabilistic by noise. Summarized through the concept of an energyprobability relationship for inverters based on AMI 0.5µm and TSMC 0.25µm processes, we quantitatively show that signiﬁcant energy savings are possible when a probabilistic inverter is switched with probability 1/2 \textless p \textless 1, and that these savings increase exponentially as p is lowered. We also quantitatively show that for a ﬁxed p, increasing the noise RMS has the effect of increasing energy dissipation quadratically. Collectively, we refer to these two facts as the energy-probability laws governing probabilistic CMOS switches—these laws constitute the ﬁrst contribution of this work. Furthermore, we also present a practical realization of a probabilistic inverter in a readily available TSMC 0.25µm technology. Finally, by using the probabilistic inverter as a building block, we provide early evidence that probabilistic switches can yield significant improvements to the energy×performance metric at the application level, by a factor of more than 288, for a probabilistic neural network application.

@article{cheemalavagu_probabilistic_nodate, title = {A {Probabilistic} {CMOS} {Switch} and its {Realization} by {Exploiting} {Noise}}, abstract = {By viewing noise as a resource rather than as an impediment, we demonstrate an entirely novel approach to ultra low-energy computing. The subject of this study is the probabilistic inverter, ubiquitous to the design of digital systems, whose behavior is rendered probabilistic by noise. Summarized through the concept of an energyprobability relationship for inverters based on AMI 0.5µm and TSMC 0.25µm processes, we quantitatively show that signiﬁcant energy savings are possible when a probabilistic inverter is switched with probability 1/2 {\textless} p {\textless} 1, and that these savings increase exponentially as p is lowered. We also quantitatively show that for a ﬁxed p, increasing the noise RMS has the effect of increasing energy dissipation quadratically. Collectively, we refer to these two facts as the energy-probability laws governing probabilistic CMOS switches—these laws constitute the ﬁrst contribution of this work. Furthermore, we also present a practical realization of a probabilistic inverter in a readily available TSMC 0.25µm technology. Finally, by using the probabilistic inverter as a building block, we provide early evidence that probabilistic switches can yield significant improvements to the energy×performance metric at the application level, by a factor of more than 288, for a probabilistic neural network application.}, language = {en}, author = {Cheemalavagu, Suresh and Korkmaz, Pinar and Palem, Krishna V and Akgul, Bilge E S and Chakrapani, Lakshmi N}, keywords = {⛔ No DOI found}, pages = {7} }

Downloads: 0