Description:

With the rapid pace of AI model testing and development and the option to run large ensembles of models to find the best combinations of model super-parameters, enhancing performance and utilization of high performance computing resources by the AI model training and testing processes can greatly reduce the time necessary to find the “best” model. Therefore, this session will look at efforts to investigate the tools and libraries used for training and evaluating AI models with an emphasis on identifying inefficiencies or introducing new algorithms that could lead to further decreases in time to solution or increase efficient utilization of HPC resources for greater scalability and higher data throughput. Examples of submissions would include but not be limited to enhanced utilization of computational resources such as CPUs and GPUs, optimal use of high performance storage and file systems to increase data throughput, fully utilizing available bandwidth of or optimizing communication patterns across interconnect networks, and finding areas to overlap communication, storage, and/or computation to reduce process idle time.