The proliferation of widely accessible graphical processing units is changing the landscape of supercomputing, offering researchers multiple benefits — and a few challenges, according to Wen-mei Hwu, Ph.D., who recently presented “The Future of Scalable Computing with GPU Computing” as an invited guest for the Ohio Supercomputer Center’s Computational Science Lecture Series.
Hwu serves as the Walter J. Sanders III-Advanced Micro Devices Endowed Chair in Electrical and Computer Engineering at the University of Illinois. OSC’s Statewide Users Group and the Ralph Regula School of Computational Science sponsored the event, held at the Wexner Center at Ohio State University and videolinked to four additional sites.
To continue improving application-level computation speeds, high performance computing will rely on parallel programming on graphical processing units (GPUs), Hwu forecast. A GPU is a specialized logic chip usually devoted to rendering 2D or 3D images.
Multiple factors are driving the parallelism revolution. In addition to the easy entrée into GPU parallel programming — naïve code typically can improve speed by two to three times — GPUs can provide sufficient numerical precision and accuracy, scale very well in the parallel domain, and are affordable and available in a range of forms, from laptops to supercomputers.
To prove his point, Hwu shared an example of an advanced MRI reconstruction. Incorporating additional information to increase image quality without increasing scan time, iterative reconstruction takes days using a traditional parallel system. However, by fine-tuning parallel programming codes on GPUs, the compute time dropped to minutes.
An exciting application of this is sodium mapping of the brain, in particular for patients with stroke or inoperable brain tumors. Typically, oncologists must wait six weeks to see if a brain tumor is responding to chemotherapy treatment, Hwu said.
“During a trial, sodium mapping with one of the highest quality MRI scanners available (9.4 Tesla) enabled doctors to see a shift in a brain tumor cells’ sodium levels – an early indicator if the prescribed treatment is working – within days,” Hwu said. “This drastic reduction in ‘time to discover’ can change way people study oncology, strokes, any behavior in the brain. We’re just beginning to test these applications.”
One of the biggest challenges of GPUs is parallel programming education. “Students can learn enough about CUDA (a GPU-accelerated application program) to write a simple parallel program in one lecture,” Hwu said. “Parallel programming is very easy, unless you care about the speed. But, it takes time to grasp the full nuances of high-performance CUDA programming.
“Industry relies increasingly on the technology of parallel processing and multi-cores,” he said. “It is vital to teach these skills to educators and students today for the United States to remain competitive.”
Through the University of Illinois Urbana-Champaign’s Institute of Advanced Computing Applications and Technology, Hwu is leading a project to develop application algorithms, programming tools, and software for the deployment of next-generation accelerators — including graphics processing units and field-programmable gate arrays — in science and engineering applications.
Ohio State University and OSC are part of the Great Lakes Consortium, which is focusing on ways to accelerate the time to discovery in science and engineering through computational science. Overall, their mission is to empower science and engineering researchers by enabling their applications to run 100 times faster and at much lower cost than on traditional parallel processing techniques. “Ohio is a partner in the consortium, and I hope that people in attendance today, especially students, will join us in this effort,” Hwu said.
The Ohio Computational Science Lecture Series is part of a joint effort by the Ohio Supercomputer Center and the Ohio Board of Regents to improve awareness and understanding of computer modeling and simulation.
Celebrating 20 years of service, the Ohio Supercomputer Center (OSC) is a catalytic partner of Ohio universities and industries that provides a reliable high performance computing and high performance networking infrastructure for a diverse statewide/regional community including education, academic research, industry, and state government. OSC promotes and stimulates computational research and education in order to act as a key enabler for the state's aspirations in advanced technology, information systems, and advanced industries. For additional information, visit http://www.osc.edu.
The Ralph Regula School of Computational Science is a statewide, virtual school focused on computational science. It is a collaborative effort of the Ohio Board of Regents, Ohio Supercomputer Center, Ohio Learning Network and Ohio's colleges and universities. The school acts as a coordinating entity for a variety of computational science education activities aimed at making education in computational science available to students across Ohio, as well as to workers seeking continuing education about this technology.