There’s overclocking, and then there’s this: University researchers are in the early stages of putting multi-core processors in smartphones to work in bursts of up to 16 times their normal speed during times of need.
In virtually all use cases, smartphones only require a fraction of their full power, but occasionally you will find yourself waiting while it chews through some hefty processing tasks. For those moments, a technique called “computational sprinting” could be the fix.
Developed by researchers at the University of Michigan and the University of Pennsylvania, computational sprinting involves briefly — for less than a second — turning on as many as 15 idle CPU cores and running in parallel. It’s even harder than it sounds, for a multitude of reasons.
“One of the challenges is it only works if there’s parallelism. We’ve been targeting image processing, one of the places where it’s comparatively easier to use multiple cores. It’s a lot harder in other domains,” said Professor Thomas Wenisch, co-author of the research paper on computational sprinting and an assistant professor at the University of Michigan’s department of Electrical Engineering and Computer Science.
This is not for playing Angry Birds or checking email, since those require sustained computation. This would be for a burst of processing, such as rendering an image or video or processing an audio signal or working out a route on a GPS map.
Cooling vs. Performance
After the burst of processing, there would be a cool down period of up to 20 seconds as the massive amount of heat generated is dissipated and more power is stored. Smartphone batteries don’t have the power to run 16 cores at once, so in order to run the sprint, power has to be stored up in a capacitor. That way, when the user runs a sprint, the power is drawn from both the battery and what’s stored in the capacitor.
It will be several years before the researchers are able to put theory into silicon, but that’s fine because the problem won’t manifest for a while. Our rush to skinnier smartphones has left handset makers in a dilemma of cooling vs. performance, and cooling won out. With more transistors being shoehorned into ARM and Atom processors, the thermal envelope is not shrinking. The result is much of the processor is switched off in order to keep the phone from overheating.
The idle transistors are dubbed “dark silicon” and it’s an increasing problem. In 2009, ARM CTO Mike Muller warned that by 2019, smartphones could be using as little as 9% of a CPU’s transistors, with the rest switched off just to keep the processor cool.
That’s because voltages aren’t scaling down at the same rate as transistors increase. So we’re getting more and more transistors in smartphone CPUs but they are not scaling down in the power they consume like they used to, so the result is power density is increasing and phones are getting hotter.
Computational Sprinting To the Rescue
Computational sprinting says go ahead, put more transistors in and leave them idle. It will use them when they are needed. Wenisch and his fellow researchers found that up to 15 additional cores could be activated at once to work in parallel alongside the chip’s main core for up to one second, which could increase the device’s response time ten-fold.
Another limitation that Prof. Wenisch and the other researchers are wrestling with is the limits of passive cooling. Heat has to bleed off somehow and they can’t make the phone any hotter than they already get. This is requiring some help from non-computer science engineers. One partner in the research project is a mechanical engineering professor who specializes in heat transfer materials.
To handle sprinting’s higher temperatures, Wenisch and team are looking at using a material that briefly melts to absorb the heat. “When you melt the material, it holds the temperature in the package at a constant level and absorbs a huge amount of heat. Over the next 30 seconds, the processor slowly cools off and the material will start cooling,” said Wenisch.
The project has a way to go. The researchers have applied for funding but can’t announce anything yet. Over the next three or four years, the group hopes to build a prototype, design the heat sink, then build the software and apps to go with the processor. Wenisch said they still have to work out how to inform the user that their phone needs to catch its breath before it can sprint again and develop software that takes advantage of sprinting.