ABSTRACT:

The re-emergence of reduced floating-point precision computer architectures has provided greater efficiencies in areas like AI model training. However, scientific computing applications where both accuracy and precision are critical have proven resistant to similar reduced-precision improvements. In this talk, we will review the implementation of a reduced-precision thermal radiative transfer code using the Implicit Monte Carlo method. Several techniques falling under the categories of arithmetic manipulations and scaling methods will be discussed for their ability to enable accurate reduced-precision calculations. Arithmetic manipulations include things such as alternate summation algorithms and automatic order of operations re-arranging, while scaling methods include both dynamic and static re-scaling of variables as well as the use of multiple energy scales. Half (16-bit) and Double (64-bit) precision results for various thermal radiation benchmark problems will be compared to demonstrate the efficacy of the reduced-precision techniques.