In a world where CPUs are no longer the be-all and end-all of computing power, MIT's Associate Professor Jonathan Ragan-Kelley is cutting through the complexity to tailor-make programming languages that amplify the capabilities of modern hardware. Ragan-Kelley, from MIT's Department of Electrical Engineering and Computer Science and a lead researcher with the MIT-IBM Watson AI Lab, is pushing boundaries to enhance how visual AI systems process graphics and images on an array of contemporary devices.
According to a report by MIT News, Ragan-Kelley is behind the Halide programming language, an instrumental tool in the industry used for photo editing and processing. "The single biggest thrust through a lot of our research is developing new programming languages that make it easier to write programs that run really efficiently on the increasingly complex hardware that is in your computer today," Ragan-Kelley told MIT News. It's not just about building new languages, but also about automating the process with smarter compilers that understand the nuances of the hardware they're talking to.
With Moore's law hitting a plateau, there's an ongoing shift from using one-size-fits-all CPUs to more specialized powerhouses like GPUs and dedicated accelerators. This pivot necessitates programming adaptability, which engineers haven’t had to grapple with before on such an intensive scale. Ragan-Kelley is pioneering a middle-ground approach with languages designed to optimize computing efficiency through a technique called user-schedulability, providing developers with safer, more efficient tools to control program performance.
"We're trying to build a new class of languages that we call user-schedulable languages that give safer and higher-level handles to control what the compiler does or control how the program is optimized," Ragan-Kelley explained in an interview with MIT News. The aim is to facilitate fine-tuning without mandating programmers resort to painstaking and potentially hazardous low-level coding.
Machine learning and AI techniques are key players in Ragan-Kelley's strategy to generate optimized scheduling interfaces for compilers, allowing for better compiler performance. He describes a groundbreaking method called "exocompilation," that brings a new perspective to compiler design, facilitating human guidance and customization for targeting specialized hardware like IBM's machine-learning accelerators. These enhancements could have a broad impact, from improving computer vision algorithms to refining speech and text generation models.
One of his notable projects involves reimagining the architecture of large language models, which could result in cheaper, more capable systems that demand less memory and fit into smaller computing devices, without losing accuracy. This work is seen as vital, forward-thinking, and uniquely positioned to reshape current computational frameworks for a future where efficiency will be paramount.
Ragan-Kelley also teaches a course on Software Performance Engineering at MIT, addressing the educational needs for understanding not just parallelism but also the optimization of memory and specialized hardware for top-tier performance in computing—tailoring the next wave of programmers to think outside the conventional programming box and embrace the potential of intricate hardware designs.
