Reconfigurable hardware could boost computer performance
A University of Wisconsin-Madison professor is studying how to use reconfigurable hardware to implement a wide range of computer accelerators that boost performance and increase energy efficiency.
The reconfigurable hardware is flexible enough to allow developers to customize the accelerators to execute multiple applications, and the research has earned Electrical and Computer Engineering Assistant Professor Katherine Compton a prestigious National Science Foundation CAREER award.
Compton says the idea of her work is similar to cooking. Chefs can make anything from a cookbook, but they can make a dish much faster if they memorize the recipe. A traditional central processing unit, or CPU, is like a chef with a cookbook; it can process anything, but it’s relatively slow since it processes data sequentially and has to look up the instructions every time, even if it's handled the same task before.
Application-specific integrated circuits, called ASICs, are hard-coded at the factory with a single “memorized” recipe. This hardware is fast because it doesn't need to look up the instructions, and it processes data in parallel, meaning it handles multiple data threads simultaneously like a graphics processing unit, or GPU.
Reconfigurable hardware is a flexible form of special-purpose hardware that is like a super chef who can quickly memorize or re-memorize a small set of recipes. Like ASICs, the hardware memorizes functions and performs computations in parallel. Reconfigurable hardware goes beyond ASICs by loading sets of data that determine which wires should be connected or disconnected, thereby creating different digital circuits for different tasks.
For example, the hardware could load an MP3 encoder accelerator to compress an audio file and then quickly switch to become a decryption accelerator.
Compton's research focuses on how a computing system determines which accelerators should be loaded into hardware at any given time. Since first publishing on system-level reconfigurable hardware management in 2005, Compton has studied how to allocate the accelerators in response to, but in isolation of, the rest of the computer system. The CAREER award, which comes with a five-year grant of more than $407,000, will allow her to expand her work to study the entire system and schedule multiple computing resources to work in tandem with the reconfigurable hardware.
In terms of the cooking metaphor, she essentially is looking at the entire kitchen workflow to determine how the various chefs—in other words, the CPU, GPU and reconfigurable hardwarecan best work together to most efficiently make the dish, or execute an application.
Ultimately, Compton is working to demonstrate to hardware companies that reconfigurable hardware provides enough of a boost to warrant adding it to everyday computing devices. “We're looking at having potentially the same or faster processing speed, but with lower energy consumption,” she says.