Grading defects: evaluating approximate circuits for error-tolerant systems

Applications such as image processing and machine learning can tolerate computational inaccuracies, allowing for the use of approximate circuits that trade precision for reduced power consumption and improved performance. Traditional testing methods, designed for exact computing, often discard a circuit upon detecting any defect, which is too strict for approximate circuits where minor deviations are acceptable. To overcome this limitation, we introduce and formalize a novel testing methodology for approximate circuits that quantifies the impact of defects on overall performance. Our method assesses the degree of approximation by measuring how defects affect computational accuracy. Instead of a binary pass/fail outcome, this technique grades the quality of the circuit, enabling a precise test-based separation of acceptable and unacceptable circuits. This approach not only enhances testing efficiency but also expands the potential for deploying approximate circuits across various fields.

Subjects

testing

approximate computing

fault tolerance

DDC Class

004: Computer Sciences

Options

Grading defects: evaluating approximate circuits for error-tolerant systems