Energy-Efficient Approximate Computation in Topaz

We present Topaz, a new task-based language for computations thatexecute on approximate computing platforms that may occasion-ally produce arbitrarily inaccurate results. The Topaz implementa-tion maps approximate tasks onto the approximate machine and in-tegrates the approximate results into the main computation, deploy-ing a novel outlier detection and reliable reexecution mechanism toprevent unacceptably inaccurate results from corrupting the overallcomputation. Topaz therefore provides the developers of approxi-mate hardware with substantial freedom in producing designs withlittle or no precision or accuracy guarantees. Experimental resultsfrom our set of benchmark applications demonstrate the effective-ness of Topaz and the Topaz implementation in enabling developersto productively exploit emerging approximate hardware platforms.