Browsing by Author "Resch, Salonik"
Now showing 1 - 2 of 2
Results Per Page
Sort Options
Item Binary Neural Networks in Spintronic Memory(2021-03) Resch, SalonikNeural networks span a wide range of applications of industrial and commercial significance. Binary neural networks (BNN) are particularly effective in trading accuracy for performance, energy efficiency or hardware/software complexity. In this thesis, I demonstrate a spintronic, re-configurable in-memory BNN accelerator, affectionately named PIMBALL: Processing In Memory BNN Accelerator, which allows for massively parallel and energy efficient computation. PIMBALL is capable of being used as a standard spintronic memory (STT-MRAM) array and a computational substrate simultaneously. For evaluation, I test PIMBALL using multiple image classifiers and a genomics kernel. Thesimulation results show that PIMBALL is more energy efficient than alternative CPU, GPU, and FPGA based implementations while delivering higher throughput.Item Novel Application Specific Architectures for Extreme Efficiency Under Harsh Operating Conditions(2022-10) Resch, SalonikLess than one hundred years ago, computation was limited to centralized and large scale machines that could barely fit in a single room, and could only be accessed by the scientists and engineers that worked on them. That world is hard to imagine today. Enabled by incredible technological advancement, and necessitated by ever increasing demand for its benefits, computing has become ubiquitous in the modern era. Modern life is filled with mobile and embedded computing devices. Additionally, the large, room sized computers have remained relevant, as powerful computers in data centers are accessible over the web, providing invaluable computational resources. However, as computing has spread into ever more domains, the challenges associated in architecting these computing systems have increased. Computers no longer simply exist in stable, controlled environments, and they must be designed to handle any and all unique challenges imposed by their surroundings. The challenges involved vary wildly by location, and include operating under extreme power restriction, tolerating intermittent power supply, avoiding corruption due to radiation, and maintaining correctness across wide temperature ranges, to only name a few. Surprisingly, these challenges are not exclusive to mobile or embedded devices. Even computers in data centers are facing new challenges, as new, more advanced devices are becoming ever more sensitive to noise which was previously of no concern. Notably, computers which can exploit quantum mechanics now must deal with new forms of noise which can not be mitigated by traditional means. Hence, even computers in the middle of data centers can be considered to exist in ``harsh'' environments, and must adapt to the conditions they impose. All of these additional challenges, along with increasing demands for performance, have caused a diversification of computer architectures. The ``one size fits all'' mindset of the microchip of the past few decades no longer universally applies. While initially these challenges seem like a burden, they have highlighted unique strengths and weaknesses of the available hardware technology and provided insight into potential new designs and computing paradigms. Hence, these new challenges are opportunities to explore alternative approaches and new architectures which may provide improved performance, energy efficiency, and robustness. These new architectures often benefit from extreme specialization, becoming exceedingly good at their niche task. Hence, these new architectures are moreso accelerators than traditional computers. This thesis covers architectures which have been designed to handle the unique challenges of their environment and the analysis of how the corresponding challenges have impacted performance, efficiency, accuracy, or total lifetime of the device. We cover architectures in a wide range of environments, from at the core of data centers, to ``beyond the edge'' batteryless devices, and even to off-earth space applications. A common theme among all the chapters in this thesis is exploiting unique proprieties of the hardware or application in order to design a system which is particularly well suited for its environment. Prominently, we have shown that processing-in-memory architectures provide a promising alternative to traditional computing systems, and that these architectures are surprisingly robust and adaptable to harsh operating conditions. We show that architectures can be built on this technology which are resilient to power outages, robust to wide temperature ranges, and nearly immune to soft errors from radiation, all while providing high performance and energy efficiency. We also explore the limitations of these architectures, such as their finite endurance and limitations of their electrical properties, and provide insight on how these limitations can be mitigated. This thesis also covers accelerators which exploit the power of quantum mechanics, which can potentially solve problems that were previously believed to be impossible. These accelerators are highly susceptible to noise, and are fragile even in the most secure environments it is possible to construct. This is currently the limiting factor and it renders them impractical for useful applications. While noise in traditional computing systems is a highly manageable and well-studied problem, these new architectures face a whole new class of noise which much harder to overcome. We analyze the impacts of this noise, and we propose strategies which can be used to mitigate it in the near term and potentially over come it in the long term.