Scheduled to be presented at "NVIDIA GTC 2023 (GPU Technology Conference)"

Zijian Xu, an engineer at Preferred Networks, will be speaking at NVIDIA GTC 2023 (GPU Technology Conference), which will be held from March 21st (Tue) to 24th (Fri), 2023.

In order to handle a large structure of 100,000 atoms using Neural Network Potential, a technique called recomputation is used to make memory usage more efficient.

For more details, please click here. You can watch the lectures and recordings by registering for free.
https://www.nvidia.com/gtc/session-catalog/?tab.catalogallsessionstab=16566177511100015Kus&search=running#/session/1666600863417001yG53

Details of the talk

Date and Time	Wednesday, March 22, 2023 12:00-12:25 JST
Speakers	Zijian Xu, Engineer, Preferred Networks
place	online
Title	Running 100,000-Atom Molecular Dynamics with Accurate NN Potential on a GPU using Automatic Recomputation
Overview of the Webinar	Accurate and large scale atomistic simulation allows a new way of material discovery. Although neural network potential (NNP), a deep learning-based force field, dramatically accelerates computation time compared to conventional DFT methods, the amount of GPU memory often becomes a bottleneck to running large-scale simulations. Supporting a larger number of atoms in such simulations greatly increases the range of structures and phenomena that can be reproduced, including large molecules like proteins, heterogeneous systems like composites, additives, and lattice-mismatched interfaces. To overcome this issue, our method analyzes and reorganizes the computation graph of the neural network model to achieve exactly the same outputs with a smaller memory budget. The basic idea here is known as “recomputation,” which wipes out less frequently referenced large data from GPU memory and recomputes them from smaller values when necessary. Our experiments have shown that it is possible to run up to 5x larger simulations with only a 25% execution time overhead using the same amount of GPU memory.