Thomas Sterling, a professor of intelligent systems engineering at the Luddy School of Informatics, Computing, and Engineering, has been awarded a NASA Phase I Small Business Innovation Research grant to develop a prototype implementation of a new class of computing accelerator.
The concepts driving the accelerator were developed over recent years by IU research sponsored by the Department of Energy, the National Science Foundation, the Defense Advanced Research Projects Agency, and the Intelligence Advanced Research Projects Activity. If successful, the development of the project will greatly speed-up forms of machine learning, neuromorphic computing, and simulation for science and engineering.
“The genesis of this particular project is the critical need for greater performance and power efficiency in spite of the severe obstacles imposed by the end of Moore’s Law, Dennard scaling, and clock-rate saturation,” Sterling said. “These had served as the foundation of exponential capability and capacity growth for the last several decades, but with the emergence of nano-scale enabling device technologies, it can no longer drive this path-setting trend. The one domain of potential improvement is in the area of computer architecture.”
A start-up company, Simultac, has been created to produce the accelerator and perform engineering tasks that are less appropriate for students and IU researchers but that are nonetheless critical to the introduction and delivery of such components to the high performance computing community. The accelerators will be plug-and-play with conventional commercial computing platforms and will augment those platforms with additional capabilities.
“For an array of application classes, exploration suggests a gain of 10X to 100X performance and efficiency gain with respect to normal multicore processor sockets,” Sterling said. “In particular, it is expected that the Simultac active memory accelerators will have dramatic effect on machine learning, dynamic graph processing, and science and engineering by speeding up simulations that use algorithms such as adaptive mesh refinement. Advanced versions of the accelerators will be seamlessly scalable (multiple accelerators working in concert) to user host processors and applications.”
The NASA SBIR program funds the research, development, and demonstration of innovative technologies that fulfill NASA needs and have significant potential for successful commercialization. Phase I of the program provides the opportunity to establish the scientific, technical, and commercial merit and feasibility of the proposed innovation with the possibility of the grant being extended to two additional phases. This phase will run through early 2020, and Sterling is joined on the project by Maciej Brodowicz, a research scientist in the Department of Intelligent Systems Engineering at Luddy.
“The type of technological innovation being pursued by Thomas and his group has the potential to push the Artificial Intelligence revolution to the next level,” said Raj Acharya, dean of the Luddy School. “Developments of this kind are a catalyst for research, and it further establishes the Luddy School as a leader in AI and machine learning.”