Two Computer Sciences faculty members, Matt Sinclair and Shivaram Venkataraman, have recently won National Science Foundation CAREER Awards. Sinclair and Venkataraman both joined the UW-Madison CS faculty in 2018 so are early (pre-tenure) in their careers. The NSF intends the awards to help early-career faculty build a firm foundation for a lifetime of leadership in integrating education and research.
Sinclair is working on designing future heterogeneous systems. He describes heterogeneous systems as systems that use “a wider variety of processing units” than CPUs. He says, “For example, your smartphone has CPUs, GPUs, and 30-50 additional accelerators. Each of these processors is optimized for different tasks – for example, GPUs are optimized for graphics.” The most common form of heterogeneous systems in modern devices are smartphones, but increasingly other kinds of systems, including supercomputers, laptops/desktops, and servers, are starting to incorporate more diverse sets of processing elements.
Sinclair says that the way we combine all these disparate processing elements is reaching an inflection point – the underlying processing techniques (based on Moore’s Law) that have been used for many years to mass produce computing products with transistors are starting to reach scaling limitations as researchers shrink the designs down to smaller and smaller parts. For example, modern designs are 5 nm, subsequent generations will be 3 nm, and so on. As transistor size continues to decrease, instead of designing one large, monolithic computer chip that combines all of the different processing units above, the community is starting to look into instead having many, small computer chips (also known as “chiplets”) and connecting them together to build larger aggregate systems. This allows hardware designers to continue scaling performance without the challenges monolithic systems are facing. This all presents a new challenge. Sinclair says, “We need to rethink how the interfaces between the different processing elements work and how to make them efficient. My work proposes a method to do this.”
Venkataraman is working on designing software systems that can support large scale data analysis and machine learning. His CAREER proposal “seeks to develop new systems that can apply Machine Learning (ML) methods on large, graph structured datasets.” Examples of large graph structured datasets included social networks, knowledge graphs in search engines and protein interaction networks. Venkataraman and his colleagues have done some early work in the development of Marius, a system that alleviates data movement bottlenecks to improve the efficiency of training ML models on graph structured data. He says, “We are currently working on systems that can make it more efficient to deploy ML models trained on graph data.”
Congratulations to Sinclair and Venkataraman on this prestigious award!