Approximate Computing

The Approximate Computing (AC) group researches means of sacrificing correctness of computations (i.e., approximating the computations) in favor of gains in non-functional aspects such as less energy consumption, less area consumption, or less latency.

The approximate computing paradigm is based on the fact that many real-world applications do not actually require the highest possible accuracy. They can, in fact, tolerate a certain amount of inaccuracy or even computational errors. The domains that render themselves well for AC are Computer Vision, Machine Learning, Multimedia, Big Data and Gaming. In these domains, approximate computations are sufficient as the human perception is limited to begin with as well as there often already is noise or redundancy in the input data.

The group investigates approximation techniques and approaches on different levels of abstraction. Results on designing specific high-performance circuits for arithmetic operations as well as on approximating whole systems for accelerating artificial neural networks have been obtained.

Current research projects

• Approximate Computing on FPGAs

 
Available resources

• Open library of approximated PLA circuits: AxPLA
• Open library of approximate adders: FAU
• Open library of approximate multipliers: AxSM
• Aarith: An Arbitrary Precision Number Library
• ABO: BDD-based Error Metric Analysis, Computation and Optimization Framework

Publikationen