Optimizing Energy Utilization of Flexible Deep Neural Network Accelerators via Cache Incorporation

There exists, to a great degree, the ever-growing desire to stress the limits of what is possible. This is especially true in an era one could ostensibly describe as “limitless advancement.” This advancement, however, comes with implicit costs. Costs which, at times, seldom make headlines. Costs which, if left unminimized or unmaintained, halt limitless advancement. For the purposes germane to this research, one can ascribe the implicit cost to energy. Specifically, the energy expenditures attributed to the installment of so-called flexible deep neural network accelerators. This body of research, then, aims to lower the energy requirements of these systems, in terms of leakage and dynamic energy, via subtle yet critical incorporation of a cache.