| Abstract: |
|
In this talk we will present our research efforts to develop a scalable machine learning toolkit that is optimized for HPC systems, and how we are starting to apply it to large scale scientific data sets. The Livermore Big Artificial Neural Network toolkit (LBANN) is a distributed memory algorithm that is optimized for high performance Infiniband networks, GPUs, node-local NVRAM, and parallel file systems. Built on the Elemental of C++ and MPI distributed linear algebra library, we use a combination of MPI+OpenMP with cuDNN and cuBLAS to create a scalable framework that is optimized for both strong and weak scaling with both model and data parallelism. LBANN has been designed to train large neural network models on very large data sets to support research projects such as the DOE-NCI BAASiC precision medicine initiative and ECP CANDLE projects. We will present preliminary efforts on training models for large scale drug discovery models and molecular dynamics simulations.
|
|
;){kind=link}