6/26/2024 Jenny Applequist
Written by Jenny Applequist
In recent years, hardware accelerators have increasingly been used to improve system performance and energy efficiency by doing specialized tasks that would otherwise be done less efficiently by CPUs. But something about accelerators has been bothering Dong Kai Wang: when applications cannot make use of them, they “just sit there doing nothing,” as he puts it.
Now, under a new grant from Samsung, Wang, who is a teaching assistant professor in Electrical and Computer Engineering, is hoping to change that. “We want to try to use the accelerator’s resources to automatically accelerate certain tasks that the CPU is currently running,” he says.
He started to consider the general problem as a UIUC Ph.D. student advised by Nam Sung Kim, who is the co-PI on the Samsung award. In his doctoral work, Wang’s approach for getting accelerators to run CPU applications was to design a whole new general processor, but he was unhappy with the results.
“It costs too much area, it’s not as flexible, and also it’s hard to convince companies to adopt this design, because... what we were proposing requires a lot of modifications. In other words, it’s very intrusive. And as much as companies like innovation, they tend to be risk-averse when you want to change too many things,” he says.
Wang therefore switched to a different strategy: rather than design an accelerator that can directly run CPU code, he decided to create a small controller that simply translates the CPU’s machine code into a format that can be run on a reconfigurable accelerator during execution.
“So this way, we essentially accomplish what we wanted, which is to run CPU code on an accelerator,” he says. “But we are also doing it in a way that’s not majorly intrusive; we’re proposing just adding this lightweight controller... that only requires minor modifications to the existing CPU and accelerator. So this way, it’s a small investment in terms of chip area... but the benefits you gain from that can be well worth the effort.”
Wang did preliminary work in which he assessed his design via simulations, but Samsung’s participation in the new project will allow him to build a real chip that incorporates the design. “This will allow us to look at, more in depth... how much actual performance is being gained, [and] efficiency improvements,” says Wang. “It allows us to basically evaluate [our translation controller] in much more detail and much more realistically.”
A robust demonstration of the solution’s capabilities could open opportunities for commercialization. Wang and Kim have already applied for a patent.