Maximizing Throughput of Power-Constrained Multi-Core Processors Manufactured with Unreliable CNFETs

Abstract

Silicon is used in making transistors conventionally but scaling of silicon if getting difficult with time.Scaling of silicon has become difficult lately.Without scaling we won?t have more transistors on chip every year for same area.Recently a trend of GPGPUs (multicores) has emerged [1] but works only for parallel applications and most applications are not parallel.Considering this researchers are looking for alternate technology which can help us get speed benefits of scaling.There are number of such alternate technologies like quantum computation in which rules of quantum mechanics will be exploited to build a fast computer,but it is very difficult to construct and there is no definite estimate when we will be able to make one.Other is carbon nanotube field effect transistor [2]. The benefits of using carbon nanotube instead of silicon is it?s high conductivity and low power consumption.It is estimated and transistor can be 5 times faster and 4 times less power consuming as compared to silicon transistor of same technology.A 20 times power delay benefit is a huge benefit but it has its problem. Carbon Nanotubes are not necessarily semiconducting like silicon. They can be metallic as well as semiconducting depending on the geometry.If a Carbon Nanotubes are manufactured without any outside interference you get around 66% semiconducting and 33% metallic,but using various manufacturing methods you can have up to 95% metallic CNTs. Our insight was instead of making CPUs which have more transistors we can have GPUs which will have many cores with less number of transistors.So the probability of error goes down and even if some cores are faulty you can just not use them.