Abstract | ||
---|---|---|
Reducing the energy consumption of todayu0027s microprocessors, for which Approximate Computing AC is a promising candidate, is an important and challenging task. AC comprises approaches to relax the accuracy of computations in order to achieve a trade-off between energy efficiency and an acceptable remaining quality of the results. A high amount of energy is consumed by memory transfers. Therefore, we present an approach in this paper that saves energy by converting data before transferring it to memory. We introduce a static approach that can reduce the energy upi¾źto a factor of 4. We evaluate different methods to get the highest possible accuracy for a given data width. Extending this approach by a dynamic selection of different storage data types improves the accuracy for a 2D Fast Fourier Transformation by two orders of magnitude compared to the static approach using 16-bit data types, while still retaining the reduction in energy consumption. First results show that such a conversion unit can be integrated in low power processors with negligible impact on the power consumption. |
Year | Venue | Field |
---|---|---|
2016 | ARCS | Computer science,Efficient energy use,Real-time computing,Fast Fourier transform,Data type,Energy consumption,Order of magnitude,Approximate computing,Power consumption,Computation |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
10 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Michael Bromberger | 1 | 4 | 2.11 |
Vincent Heuveline | 2 | 179 | 30.51 |
Wolfgang Karl | 3 | 46 | 6.44 |