Towards Practical Stochastic Computing Architectures for Emerging Applications

Thesis (Ph.D.)--University of Washington, 2019The end of Dennard scaling and demands for energy efficient, low power, and high density computing solutions over the past decade has forced exploration of new computing technologies. Stochastic computing is one of these alternative computing technologies which has enjoyed renewed interest and is the primary focus of this dissertation. Stochastic computing is a form of approximate computing which encodes values as probabilistic bitstreams where the ratio of 1s and 0s determines the encoded value. This representation allows stochastic computing to achieve lower operating power, higher computational density, and better error resilience compared to conventional binary-encoded circuits. In its current form, stochastic computing presents a number of challenges before it can become a practical replacement for conventional binary-encoded computing. First, there is little prior work detailing design methodologies to guide effective implementation and integration of stochastic computing into accelerator architectures. Second, the application space where stochastic computing yields compelling gains is far from obvious and has only seen limited exploration. Third, stochastic arithmetic circuits are unintuitive to design because they require careful consideration of correlation and quantization effects. This thesis focuses on new circuit components, applications, architectural considerations, and design techniques to improve the practicality of stochastic computing accelerators. I first propose novel stochastic circuits to improve the accuracy of stochastic computations and augment the range of implementable functions. I then evaluate the viability of stochastic computing with a design space exploration of end-to-end stochastic computing accelerator architectures. In this exploration, I evaluate under what design parameters and conditions stochastic computing accelerators are competitive alternatives to their binary-encoded counterparts. Using these guidelines, I use these results to establish a set of architecture design guidelines to help designers identify when and why they should consider stochastic computing. I then evaluate codesign opportunities and empirically measuring power, area, and energy efficiency for emerging applications. I also propose borrowing techniques from program synthesis such as stochastic synthesis and mixed integer linear programming to automatically synthesize novel stochastic circuits. Finally, I conclude with future directions for further improving the practicality of stochastic computing as well as additional research directions beyond stochastic computing