Measuring the gap between programmable and fixed-function accelerators: A case study on speech recognition

Abstract

As power and energy consumption have become the key design constraint of mobile systems, mobile system-on-chip (SoC) architects have dedicated a progressively larger area budget to custom accelerators: graphics processors, audio/video codecs, and image signal processors abound. Fixed-function accelerators now occupy more than half of the die area of these chips [2], and we foresee this trend only to progress in the near future. Indeed, for applications that have adopted a well-defined standard, constructing dedicated fixed-function accelerators may prove to be the best approach [5]. But for applications that are not standardized and are still in flux, we ask: Is building fixed-function accelerators the right answer, or should dedicated accelerators become more programmable? To this end, we first measure the performance and energy consumption gap between programmable and fixed-function accelerators for automatic speech recognition (ASR), an application that is not yet standardized but whose adoption is on on the near horizon. We then analyze these results to gain insight into this problem.