GTC-DC 2019: On the Merger of High-Performance Computing and Machine Learning Within Computational Fluid Dynamics
Note: Viewing this video may require joining the NVIDIA Developer Program or login in
GTC-DC 2019: On the Merger of High-Performance Computing and Machine Learning Within Computational Fluid Dynamics
Dirk Van Essendelft, The National Energy Technology Laboratory
We’ll cover recent developments on the merger of high-performance computing and machine learning (ML) within computational fluid dynamics at the National Energy Technology laboratory in collaboration with NVIDIA. We’ll address three main topics: the use of machine learning frameworks (TensorFlow) for traditional HPC applications; performance benchmarks comparing traditional HPC methods with the same methods in TensorFlow on a variety of hardware; and integration of machine learning within the framework to accelerate computations. The replacement of CFD solvers with machine learning accelerated solvers in configurations which control tolerances showed that significant accelerations are possible. While this is specific to CFD, the principles learned can easily be adopted by many other branches of engineering and science to achieve similar accelerations.