Similar Projects

Before we started with our project, we had a look on previous efforts in the field of approximate computing. In this section we introduce similar projects to ours, that we know of, that built approximate processors. Three different approaches that achieved considerable speed-up or energy savings are introduced below:

Truffle was designed by a team from the University of Washington to exploit the concept of voltage scaling. It uses two pipelines supplied with two different levels of voltage for precise and approximate processing. Execution of approximate arithmetic operation using a lower voltage supply resulted in more than 50% energy reduction for certain testbenches.

Quora a project from Purdue University, uses precision scaling with error monitoring and compensation to facilitate quality programmable execution in a vector architecture. LSBs of the operands of arithmetic operations are neglected according to Quality Programmable Instructions, wherein programmers define average or maximum error allowed. Logic slices responsible for operating on the neglected bits are then clock gated. Despite the overhead introduced by the control and monitoring logic, The approximation achieved more than 2.5X savings for 7.5 % quality loss.

A third approach uses machine learning-based transformations to accelerate approximation-tolerant programs. Certain code regions are transformed from a von Neumann model to a neural model. A neural processing unit (NPU) is tightly coupled to the processor pipeline to permit profitable acceleration even when small regions of code are transformed. For a set of diverse applications, NPU acceleration provides whole-application speedup of 2.3X and energy savings of 3.0X on average with average quality loss of utmost 9.6%.

PACO followed a similar approach to Quora to approximate addition, subtraction and multiplication operations. In addition, we adopted a rather different technique, hardware lookup tables, to approximate arithmetic functions.