Abstract Details
Abstracts
Author: M Cianciosa
Requested Type: Poster
Submitted: 2024-04-01 10:37:46
Co-authors: D. Batchelor, W. Elwasif
Contact Info:
Oak Ridge National Laboratory
1 Bethel Valley Rd
Oak Ridge , TN 37830
United States
Abstract Text:
Determining the optimal placement and orientation of RF heating when optimizing a fusion pilot plant requires exploring multiple configurations. Graphics Processing Units (GPUs) are optimized for computing many embarrassingly parallel problems. Problems where the same physics is applied to many independent calculations. However, vendor incompatibility, data locality, and architecture of legacy codes make exploiting the power of GPUs challenging for physicists.
We present a new RF Ray tracing code built on a graph computation framework. This framework is designed to enable the physicist to write the equations that are automatically translated into optimized GPU kernels. The framework builds a graph data structure representing the equations of interest. In graph form, algebraic simplification is applied to reduce the problem. Auto differentiation is enabled by applying the chain rule at each operation of the graph. The graph can then be Just-In-Time compiled to an optimized kernel that can be transparently deployed to Nvidia, AMD, and Apple GPUs or any CPU system. Using this framework, we can demonstrate fast power absorption calculations in realistic tokamak and stellarator equilibria at a massive scale.
Work supported by U.S. DOE under Contract DE-AC05-00OR22725 with UT-Battelle, LLC
Comments: