Scientific Highlights NES

Highlight DIRK 02/2023 teaser

Approximate Computing for Nuclear Reactor Simulations

During the last decades, computing power has been subject to tremendous progress due to the shrinking of transistor size as predicted by Moore’s law. However, as we approach the physical limits of this scaling, alternative techniques have to be deployed to increase computing performance. In this regard, the next big advance is envisioned to be the usage of approximate computing hardware based on field-programmable gate arrays and/or digital-analogue in-memory circuits. Such approximate computing can provide disproportional gain (x1000) in energy efficiency and/or execution time for acceptable loss of simulation accuracy. This could be highly beneficial in order to accelerate computational intensive simulations such as reactor core analyses with higher resolution multi-physics models. On the other hand, the execution of programming codes on low-precision hardware may result in inadequate outcomes due to quality degradation and/or algorithm divergence. To address these questions, studies on the stability and the performance of advanced reactor simulation algorithms as function of reduced floating-point arithmetic precision are being conducted at the laboratory for reactor physics and thermal-hydraulics. Results obtained so far indicate a large room for the acceleration of nuclear engineering applications using mixed-precision hardware. Therefore, research is now being enlarged towards assessing multiprecision computing methods for reactor core simulations with higher spatial resolution.

Read more