Enabling New Computation Paradigms with Emerging Technologies
Open Access
- Author:
- Tsai, Wei Yu
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 06, 2017
- Committee Members:
- Vijaykrishnan Narayanan, Dissertation Advisor/Co-Advisor
Vijaykrishnan Narayanan, Committee Chair/Co-Chair
John Morgan Sampson, Committee Member
Mary Jane Irwin, Committee Member
Bruce Gluckman, Outside Member - Keywords:
- Non-Boolean Computation
Neural Networks
Coupled Scillators
Low-power Computation
Image Processing
Audio Processing - Abstract:
- For the last decade, Moore's law has slowed down while Dennard scaling has come to an end. Thus, performance improvements can no longer rely on the shrinking of transistor sizes. On top of that heat-sink and power budgets limit the computation per unit chip area. Therefore, new technologies are being invented to go beyond the scope, and break the limitations, of traditional computer architecture using Boolean operations. Computing physics, which deploy computations onto artificial components with configurable behaviors using physical phenomenon, is one category of alternative possibilities. For example, HyperFET coupled-oscillators can be used to find the matching degree efficiently, utilizing physical phenomenon along with the electrical current controlled by Boolean logic. In addition, brain-inspired neuromorphic hardware can be built to perform the inference of the neural networks that classify and make decisions in a fault-tolerant manner. While these new technologies show great potential in performing the tasks with native support, this work explores new computation paradigms that can further improve the functionality and energy efficiency of the systems. These new computation paradigms include configurable coupled oscillators, supported by thorough modeling from device-, circuit-, to system-level, and the approaches to map the data-preprocessing onto a digital emulation platform of spiking neural networks with either explicit-programming or parrot-training. This work not only shows the successful implementations of the above and their power benefits, but also provides insights to designing systems using these new computing technologies.