Enabling Accelerator Centric Computing

With power limitations imposing hard bounds on the amount of a chip that can be powered simultaneously, but advances in manufacturing technologies continuing to pay dividends in terms of feature density, together leading to the presence of dark silicon, it becomes clear that continued advances in performance will come in the form of energy efficiency and customization rather than scaling processor count and cache size. This observation is the basis for the argument that accelerators, highly customized logic blocks that perform a particular task with both high performance and energy efficiency, are going to become increasingly relevant in future processors. It is predicted that the number of these accelerators will exceed 1500 by 2022.As accelerators become more responsible for shouldering a greater portion of computation, it becomes important to elevate accelerators to be considered a first-class computational primitive, rather than an unusual device that requires extraordinary measures to interact with. Simply having a powerful compute engine in a machine is meaningless if it is impossible to efficiently communicate with it, or if software that uses an accelerator is hard to write, or if interacting with the device involves complicated performance considerations which make it difficult to predict whether any benefit would be had by using the accelerator.The work described herein attempts to address this issue, and providing architectural extensions that allow for accelerators to become a high performance, highly efficient, and highly utilized compute elements. The effort comes from two directions: 1) introducing enabling technologies that allow accelerators to be efficiently used by software, and 2) redesigning system components to allow for accelerators to perform well and leverage existing system resources efficiently. This results in accelerator-centric designs, where conventional processing cores act more as choreographers for a communicating network of accelerators as opposed to cores acting as the primary mechanism of performing computation. The intent is to accomplish this in such a way as to place undue burden on application programmers, by introducing accelerators in such a way as to be compiler-friendly