Tiger Team: tt-ml-gpus
Quantum Dynamics from Artificial Neural Networks
The goal of this project was to support a user to make a swift transition from CPUs to GPUs. In 2020, monitoring of JUSTUS 2 revealed that a user performed machine learning research on CPUs using a software framework written in Python. Since neural networks are parallel, GPUs are far better suited than CPUs to perform numerical calculations on such objects. The user was contacted and a roadmap to port his Python code from CPU to GPU was jointly established. Detailed instructions about how to run batch jobs on nodes equipped with GPU accelerators were given to the user and the preexisting CUDA programming environment on JUSTUS 2 was extended with the installation of cuDNN. cuDNN is a GPU-accelerated library of primitives for deep neural networks from NVIDIA. Furthermore, examples about how to run JAX for Python efficiently on GPUs was provided to the user. Using these informations, the user was able to adopt the examples and successfully port his Python code to GPUs. According to the user, the performance of the code was enhanced by a factor of 3, enabling him to perform significantly more simulations and/or to process larger neural networks within the same instance of time.
Members of the Tiger-Team:
Kirchhoff-Institut für Physik, Universität Heidelberg; HPC-Kompetenzzentrum für computergestützte Chemie und Quantenwissenschaften, Universität Ulm