for Multimedia and Machine Learning
The concept of error tolerance, i.e. accepting error in a design to save resources, is well known in many abstraction layers. Built on these ideas, approximate computing has emerged as a promising candidate to improve performance and energy efficiency beyond technology scaling.
Designing approximate circuits explores a new trade off, not only by accepting unreliability, but by intentionally introducing errors to save area and power and overcome the limitations of traditional circuit design.
With the exploding amount of data being processed in the cloud and on mobile devices, a wide range of applications can trade accuracy without compromising functionality or user experience. In multimedia applications, a small proportion of errors stays imperceptible to humans. Machine-learning applications have also been proven extremely tolerant to approximation.
The ICLAB interests on approximate systems cover several research directions:
- Approximate arithmetic architectures
- Voltage/frequency over-scaled circuits
- Reconfigurable hardware