Adaptive Distributed Architectures for Future Semiconductor Technologies.

Year after year semiconductor manufacturing has been able to integrate more components in a single computer chip. These improvements have been possible through systematic shrinking in the size of its basic computational element, the transistor. This trend has allowed computers to progressively become faster, more efficient and less expensive. As this trend continues, experts foresee that current computer designs will face new challenges, in utilizing the minuscule devices made available by future semiconductor technologies. Today's microprocessor designs are not fit to overcome these challenges, since they are constrained by their inability to handle component failures by their lack of adaptability to a wide range of custom modules optimized for specific applications and by their limited design modularity. The focus of this thesis is to develop original computer architectures, that can not only survive these new challenges, but also leverage the vast number of transistors available to unlock better performance and efficiency. The work explores and evaluates new software and hardware techniques to enable the development of novel adaptive and modular computer designs. The thesis first explores an infrastructure to quantitatively assess the fallacies of current systems and their inadequacy to operate on unreliable silicon. In light of these findings, specific solutions are then proposed to strengthen digital system architectures, both through hardware and software techniques. The thesis culminates with the proposal of a radically new architecture design that can fully adapt dynamically to operate on the hardware resources available on chip, however limited or abundant those may be.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/102405/1/apellegr_1.pd