Neural Acceleration for General-Purpose Approximate Programs

Abstract

This paper describes a learning-based approach to the acceleration of approximate programs. We describe the Parrot transformation, a program transformation that selects and trains a neural network to mimic a region of imperative code. 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 accelerate small code regions. Since neural networks produce inherently approximate results, we define a programming model that allows programmers to identify approximable code regions-code that can produce imprecise but acceptable results. 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 wholeapplication speedup of 2.3× and energy savings of 3.0× on average with quality loss of at most 9.6%.

Quantifying sources of error in McPAT and potential impacts on architectural studies

“…One possible reason is that many past studies have used cores that did not trigger some errors we observed. For example, older and simpler cores like Atom and Penryn have lower issue widths, so studies using them [13,18,42,51,55] avoid read/write port overestimates, one of the major error sources we observed. Penryn cores also do not support SMT, eliminating the duplication of hardware error.…”

Exploring architectural heterogeneity in intelligent vision systems

“…Recently, there has been significant interest in non-traditional acceleration platforms, such as approximate computing designs, some of which have targeted the computer vision space. In [13], a dynamically trainable neural network accelerator is trained on regions of image processing algorithms and replicates their output with a high degree of accuracy and significant efficiency gains compared with a general purpose processor. In [30] voltage-coupled oscillators are used to build pattern-matching operators that in turn perform image processing tasks.…”

BRAINIAC: Bringing reliable accuracy into neurally-implemented approximate computing

“…First, we note that many high performance applications, such as the "Recognition, Mining, and Synthesis (RMS)" workload from Intel [21], are based on heuristics that can be approximated (e.g. with the use of neural networks [24]). Second, applications that use exact computation also often include regions of computation that are tolerant to imprecision, or "approximable", even if these regions can only be circumstantially approximated (e.g.…”