Christopher Batten almost left the field of computer architecture.
"After a couple of years in graduate school at MIT, I started to become disillusioned with my research community. It seemed like most of the academic research involved small, incremental paper designs, not the revolutionary machines I had learned about as an undergrad. I enjoyed the work I was doing on low-power parallel processors, but industry and the rest of academia were less interested. I spent a year working on synthetic biology, and loved the fact that the field was brand new and essentially wide open."
Just when Batten had to make up his mind, manufacturers turned to multi-core processors to enable more powerful chips that wouldn't overheat. "All of a sudden things became really interesting, really fast, and much of the work we were already doing was directly applicable to these new challenges," said Batten. "David Patterson from U.C. Berkeley likes to say that the industry made a Hail Mary pass and the football is still in the air. And what's fun is both industry and academia are running as hard as they can to figure out how we can catch this."
Batten is currently interested in combining the advantages of graphics processors (GPUs) and general-purpose processors in a single chip. As GPUs have become more powerful, programmers have started running applications on them that have nothing to do with graphics. "If you run a specific type of application on a GPU, you will get great efficiency and performance," said Batten, who does some of his best thinking hiking through Cascadilla Gorge on his way to work. "But if you want to do something a little different it might not be possible or the performance might be dismal."
"There's been talk of a convergence between graphics and general-purpose processors", said Batten, "But what people propose is a Frankenstein–taking standard GPU and general-purpose processors and integrating them onto the same chip," he said. "What I'm trying to do is create a true GPDPU, general-purpose data-parallel unit. I want to tightly integrate these two pieces so that it's one elegant whole, not two pieces bolted together." Batten's goal is to go to industry and make a case for this architecture to be used in the future. His hope is that it will get great performance on data-parallel programs, be able to run arbitrary general-purpose programs, and be energy efficient while doing both.
In addition to his research on energy-efficient parallel architectures, Batten and Professor Rajit Manohar are advising the newly revised Big Red Chip student project team. "The goal is to involve undergrads and M.Engs and challenge them to take what they're learning in their classes and apply it to real chip design," he said. "We hope to build up the design over the next two years, tape-out a chip in the third year, and then in the fourth year we'll get it back and do the testing and evaluation. Undergraduates at other institutions would never really get a chance to do this, so a project like this is a little crazy but also very exciting."