Reducing data movement with approximate computing techniques

Data movement is the dominant factor that limits performance and efficiency in today's architectures, and we do not expect that to change in future architectures. In this paper, we describe how approximate computing techniques can be applied to communication at the algorithm level, in conventional computer architectures, and in the architectures being explored as we go beyond Moore's Law. We present results that demonstrate potential performance gains and the effect of approximations in traditional computer architectures. We describe how these techniques may be applied to future architectures based on probabilistic, approximate, stochastic, and neuromorphic computing, as well as more conventional heterogeneous and 3D architectures.