When you shrink electronics and their transistors with them, they become more error-prone. Usually, that’s compensated for by a boost in voltage — but researchers at MIT think we should embrace the mistakes instead.
The idea is simple. Smaller chips with increasingly small transistors are more efficient, but that efficiency is squandered by pumping in power to retain accuracy. Instead, maybe we could allow chips to get some of their calculations wrong and take advantage of the possible energy savings.
You’d have to be careful where the errors occurred, of course. The odd mishap during a video render would probably go unnoticed forever, while calculations in a nuclear power plant probably need to be 100 per cent correct. But now, MIT researchers have developed a technique which lets them work out which calculations they can assign to unreliable chips — to make the most of efficiency saving without botching anything critical.
The new tool’s called Chisel and it automatically tags operations with accuracy requirements — which means they can be passed on to the appropriate piece of hardware. Tested on a series of image-processing and financial-analysis algorithms, it appears to be able to save between 9 to 19 per cent of power.
That a pretty compelling stat. It’s not hard to imagine hardware coming equipped with a series of processing cores that can be used to compute important operations with a high-spec processor and less important ones with a lower-spec, more inaccurate chip. The future might be a little bit broken — and all the more efficient for it. [MIT]