Approximate Computing: Making Mobile Systems More Efficient

摘要: Approximate systems can reclaim energy that's currently lost to the "correctness tax" imposed by traditional safety margins designed prevent worst-case scenarios. Researchers at University of Washington have co-designed programming language extensions, a compiler, and hardware co-processor support approximate acceleration. Their end-to-end system includes two building blocks. First, new programmer-guided compiler framework transforms programs use approximation in controlled way. An C Compiler for Energy Performance Tradeoffs (Accept) uses programmer annotations, static analysis, dynamic profiling find parts program that are amenable approximation. Second, targets on chip (SoC) augmented with efficiently evaluate coarse regions code. A Systolic Neural Network Accelerator Programmable logic (Snnap) is accelerator prototype code general-purpose program.