As improvements in per-transistor speed and energy efficiency diminish, radical departures from conventional approaches are needed to continue improvements in the performance and energy efficiency of general-purpose processors. One such departure is approximate computing, where error in computation is acceptable and the traditional robust digital abstraction of near-perfect accuracy is relaxed. Conventional techniques in energy-efficient computing navigate a design space defined by the two dimensions of performance and energy, and traditionally trade one for the other. General-purpose approximate computing explores a third dimension—error—and trades the accuracy of computation for gains in both energy and performance. Techniques to harvest large savings from small errors have proven elusive. This paper describes a new approach that uses machine learning-based transformations to accelerate approximation-tolerant programs. The core idea is to train a learning model how an approximable region of code—code that can produce imprecise but acceptable results—behaves and replace the original code region with an efficient computation of the learned model. We use neural networks to learn code behavior and approximate it. We describe the Parrot algorithmic transformation, which leverages a simple programmer annotation ("approximable") to transform a code region from a von Neumann model to a neural model. After the learning phase, the compiler replaces the original code with an invocation of a low-power accelerator called a neural processing unit (NPU). The NPU is tightly coupled to the processor pipeline to permit profitable acceleration even when small regions of code are transformed. Offloading approximable code regions to NPUs is faster and more energy efficient than executing the original code. For a set of diverse applications, NPU acceleration provides whole-application speedup of 2.3× and energy savings of 3.0× on average with average quality loss of at most 9.6%. NPUs form a new class of accelerators and show that significant gains in both performance and efficiency are achievable when the traditional abstraction of near-perfect accuracy is relaxed in general-purpose computing.
It is widely understood that energy efficiency now fundamentally limits microprocessor performance gains. CMOS scaling is no longer providing gains in efficiency commensurate with transistor density increases.7, 15 As a result, both the semiconductor industry and the research community are increasingly focusing on specialized accelerators, which can provide large gains in efficiency and performance by restricting the workloads that benefit. Recent work has quantified three orders of magnitude of difference in efficiency between general-purpose processors and ASICs.14 The community is facing an "iron triangle" in this respect; we can choose any two of performance, energy efficiency, and generality at the expense of the third. Before the traditional trend of transistor scaling—Dennard scaling5—broke down, we were able to improve all three on a consistent basis for decades. In this post Dennard scaling era, solutions that improve performance and efficiency while retaining as much generality as possible are highly desirable; hence the exploding interest in GPGPUs and FPGAs. Such programmable accelerators exploit some characteristic of an application domain to achieve efficiency gains at the cost of generality. FPGAs, for example, exploit copious, fine-grained, and irregular parallelism while GPUs exploit many threads and data-level SIMD-style parallelism. Whether an application can use an accelerator effectively depends on the degree to which it exhibits the accelerator's required characteristics.