Two new GPU servers recently acquired by the American Family Insurance Data Science Institute (DSI) at UW-Madison honor the memory of Dr. Olvi Mangasarian, who passed away in March 2020. Dr. Mangasarian, John von Neumann Professor Emeritus of Mathematics and Computer Sciences at UW-Madison, was a pioneer and leader in the field of mathematical programming.
“Dr. Mangasarian’s fundamental contributions range from abstract theory to practical applications,” says Glenn Fung, a data science and artificial intelligence expert at American Family Insurance who was a student of Dr. Mangasarian. “His results have been characterized as very elegant, having great impact and providing the basis for many recent advances in the field of machine learning.”
The two fully equipped 8xA100 NViDIA GPU servers, named Olvi-1 and Olvi-2, are a dedicated campus resource for data science research. The servers are connected to the UW high-speed network, with access to other UW computation and data storage resources. While there are many GPUs in use on campus, these top-of-the-line units will allow researchers to address computing bottlenecks by running larger jobs.
“A GPU, or graphics processing unit, is a specialized processor that is really good at computations for graphics,” says David Parter, Director of Academic and Computing Services in the UW-Madison Computer Sciences Department. “Being specialized for graphics makes it really good at other specialized computations, like deep learning.”
Artificial intelligence (AI) and machine learning are behind many recent, innovative applications in both academia and industry. A subset of machine learning—deep learning—mimics the activity of neurons in the human brain to speed up the process of training machines to do tasks like recognize objects and human language. Deep learning has real-world applications in areas ranging from language interpretation to healthcare to customer service.
Deep learning can have a significant computational footprint, however, and GPUs can dramatically speed up computational processes. While GPUs are in high demand and are expensive, the alternative for researchers is to rent them in the cloud. In many cases, this isn’t a cost-effective solution.
Vikas Singh, Professor of Biostatistics and Medical Informatics in the School of Medicine and Public Health, and Zhanpeng Zeng, Computer Sciences graduate student in the School of Computer, Data and Information Sciences, are using one of the new GPU servers for research supported by the DSI’s American Family Funding Initiative. Their work focuses on efficient training of transformer models for use in natural language processing and object recognition for images. Without access to the GPU server, they might have to wait many weeks and spend thousands of dollars for cloud services.
“This resource will significantly strengthen our collaboration with American Family,” says Singh. “It directly impacts the quality of the results we are putting out, and it potentially helps the university establish new research collaborations.”
The UW-Madison Computer Systems Lab is administering the GPU servers. DSI researchers, including those working on projects funded through the American Family Funding Initiative, have priority for GPU server use. Computational cycles not used by DSI researchers and affiliates will be made available to campus researchers in an inclusive and equitable manner.
For more information, contact Steve Goldstein, DSI Program Manager: firstname.lastname@example.org.