Cray Boosts Deep Learning Performance for Geospatial AI

Print Friendly, PDF & Email

Today Cray announced enhanced capabilities to empower data scientists and engineers who are innovating in the field of Geospatial AI. Cray introduced a new Geospatial Reference Configuration as well as new features in its Cray Urika-CS and Urika-XC AI and Analytics software suites.

The new features include an augmented Deep Learning Plugin that provides best-in-class deep neural network performance training and broadened support for deep learning frameworks. In performance studies, the plugin showed training time reductions up to 23% over open source alternatives for a single node, dense GPU configuration.

Both the reference configuration and plugin are designed for IT and AI teams implementing complex infrastructure to support Geospatial AI workloads. Cray also announced that it has delivered and installed a Cray CS Series system at the U.S. Geological Survey agency to support AI initiatives in geospatial analysis and the agency’s mission to provide reliable information for understanding the Earth.

Geospatial AI is the marriage of geospatial data and artificial intelligence. It promises to be one of the most important uses of AI across a range of industries such as oil and gas companies, state and local governments, property and casualty insurance businesses, weather forecasting centers, and beyond. Data scientists are exploring the use of AI, deep learning and machine learning to deliver new applications and insights based on geospatial data. For example:

Oil and gas companies will perform market supply analysis by applying AI to satellite images of tank farms and refineries.
Municipal governments will use AI to detect changes in satellite imagery for infrastructure planning and disaster and resiliency response planning.
Property and casualty insurance businesses will apply AI to satellite imagery for disaster impact analysis and claim fraud detection.
Weather forecasters will make more accurate predictions because Geospatial AI uncovers new information, such as soil moisture, with high resolution.
Shorter Training Times Advance Geospatial AI Innovation.

As Geospatial AI becomes core to organizational missions, the time to develop and refine neural network models at optimal accuracy becomes a challenging factor to innovation. To shorten the time data scientists spend developing Geospatial AI applications, Cray is releasing updates to the Urika-CS and Urika-XC AI and Analytics software suites. The augmented Cray Programming Environment (PE) Deep Learning Plugin will significantly reduce training times for complex neural network models. Internal performance studies, using the widely-available ResNet-152 and Inception-V4 neural network models, have shown significant training time improvements. Coupled with Cray’s hyperparameter optimization capabilities, the Cray Urika AI and Analytics suites dramatically improve data scientist productivity and accelerate the development of advanced Geospatial AI applications.

New Reference Configuration for Geospatial AI

The availability of new sources of geospatial data is driving the adoption of AI. Implementing a Geospatial AI workflow requires a balanced system that is able to handle the demands of data preparation and model development. Cray is introducing a new Geospatial AI Reference Configuration comprised of CS-Storm 500NX GPU accelerated nodes and CS500 CPU nodes that will be able to handle the entire Geospatial AI workflow.

Geospatial AI presents both data and compute challenges for data science and IT teams tasked with developing new applications. Our forte has long been understanding performance issues and improving performance with supercomputing technologies,” said Per Nyberg, vice president market development, AI at Cray. “Complete systems optimized for the geospatial workflow and enhanced with high-performance deep learning eliminate boundaries faced by geospatial teams exploring and implementing advanced AI applications.”

USGS Chooses Cray for Geospatial Innovation

The US Geological Survey (USGS), the science arm of the U.S. Department of the Interior, has selected a Cray CS Series system to further the use of AI in natural sciences. USGS is active in promoting the use of machine and deep learning in areas ranging from earth observation, numerical weather prediction, hydrology, solid earth geoscience and land imaging.

The updated versions of the Urika-CS AI and Analytics software suites and the Geospatial Reference Configuration are expected to be available within 30 days.

To learn more and to see a live demo of Cray geospatial capabilities, stop by the Cray booth #E-921 at ISC 2019.

Sign up for our insideHPC Newsletter