Traditional computing systems and applications have been designed under the conservative assumption that each calculation must be accurate. This design principle held so far only thanks to the safe path to computational efficiency assured by Moore’s law and its constant exponential improvement. This era is unfortunately at its twilight: from tiny nodes of the Internet-of-Things (IoT), to large computing centers of High Performance Computing (HPC), sub picoJoule/operation energy efficiency is essential for practical realizations. To overcome the “power wall” a shift from traditional computing paradigms is now mandatory.
OPRECOMP aims to build the foundation for computing based on transprecision analytics. The nature-inspired driving principle behind the approach is that almost any application involves a large amount of intermediate calculations, whose accuracy is irrelevant to the final user, who is interested only in the reliability and validity of the final result. The main reason why in today’s systems these intermediate steps are still performed in full-precision is that existing hardware, runtime software, programming environments and algorithmic approaches do not offer well-assessed solutions to relax accuracy in a controllable way without sacrificing reliability. Lifting this limitation allows a slew of opportunities to arise, which will reduce time- and energy-to-solution by orders of magnitude, without compromising the accuracy and reliability of the final result.
The “disruptive” mission of OPRECOMP is to demonstrate that this idea holds in a huge range of application scenarios: from the sub-milliWatt to the MegaWatt range, spanning nine order of magnitude. The project will build a complete framework for transprecision computing, covering devices, circuits and architecture, software tools, algorithms and data structures along with the mathematical theory and physical foundations of the ideas; OPRECOMP will demonstrate that transprecision computing is the way to think about future systems, and that the impact of the intermediate approximations in the final result can be bounded and controlled. Throughout the project a series of working platforms will be used as demonstrators and for verification in view of a subsequent industrial exploitation. The quality of transprecision computing will be proved on different applications from the mW range in the IoT domain, up to the MW domain in Big Data Analytics, Deep Learning and HPC simulations
OPRECOMP plans to develop a radically new computing paradigm and ecosystem. Inspired by nature and intelligence of human brain the developed systems are designed to spend just the right amount of energy that is required for performing any particular operation. While for financial applications the proposed systems will be adopting high precision arithmetic others such as data-mining or human-consuming applications will have the possibility to adopt less precise arithmetic and even embrace errors incurred by underpowered or new emerging circuit technologies.