Over the past year, MSI has been increasing resources available to researchers working in machine learning fields. Deep learning methods have been getting a great deal of attention due to recent advances in hardware and successful applications in the fields of image classification, computer vision, robotics, and natural language processing. In addition to consulting for specific projects, MSI has installed software and created tutorials to assist researchers getting started with machine learning techniques.
MSI has installed software including Caffe, Tensorflow, Keras, and Horovod to allow researchers to take advantage of GPU resources as well as parallel scaling across multiple nodes. In particular, Horovod has demonstrated very effective scaling for Tensorflow workloads across multiple GPU nodes at MSI. Below is a graph of the times for an MNIST benchmark on 1–64 GPUs on MSI’s k40 nodes in the Mesabi cluster.
MSI staff continues to test new environments and hardware to make available to University researchers. MSI started a Deep Learning user forum to facilitate efforts on deep/machine learning at the University of Minnesota use HPC resources. To stay in touch with new machine learning resources at MSI, please join the Deep Learning Google Group.
posted on July 16, 2018