Connect with us

Hi, what are you looking for?

Tech & Science

Google uses machine learning to make computers faster as they age

The processing speeds of modern computers are far higher than the transfer rates of the memory modules they connect to. This means that fetching data from memory incurs a performance penalty and leaves the processor idling for a moment. To avoid this, processors prefetch data ahead of time, using predictions to guess what information’s likely to be needed next.
This week, Google published a paper which explains how prefetching could be significantly enhanced using machine learning. Although the company’s not provided any demonstrations of speed increases, the MIT Technology Review notes the system could offer a considerable real-world performance boost. Machine learning lends itself well to scenarios that require iterative refinement of predictive algorithms, so prefetching may be an ideal target for the technology.
The project is part of a larger Google effort to enhance computer hardware using AI. The team behind the project said it could be used to optimise every computer component. Intelligence will enable hardware and software to integrate more seamlessly, while giving the overall system the ability to anticipate its user’s actions.
READ NEXT: Google publishes new research into how neural networks “think”
Using machine learning, a computer could direct resources to different components as the user interacts with on-screen controls. This process would be seamless and dynamic, with the hardware automatically fine-tuning itself to suit each new task. While this tech is still some way off – the demanding hardware requirements of current-generation AI makes it prohibitive on existing machines – Google’s work to improve memory prefetching provides a glimpse into the future.
According to Google, AI-powered prefetching offers “significantly higher” precision than conventional implementations of the tech. However, it also introduces its own issues. A highly optimised AI might fetch data too early, which could cause information already in the processor’s cache to be evicted prematurely, before work has been done on it. The researchers are now looking at ways to predict more steps in advance, which would help to mitigate the problem of early prefetches.
With advances in hardware performance becoming more difficult to achieve, machine learning could provide a new way to scale computer systems. Hardware that automatically tunes itself could help machines to adapt to specific scenarios, without necessarily demanding changes to the physical components. Google said its research opens “a wide range of exciting directions” for further investigation of the subject.

Written By

You may also like:

Tech & Science

Facebook and Instagram parent company Meta has said it will lay an undersea cable stretching across five continents to carry data.

Business

Data is your GPS — but are you even setting a destination?

Life

Some of these symbionts discovered are close relatives of bacteria pathogenic to animals, including humans.

Life

What’s different now is that more children are presenting with co-infections, such as RSV and flu simultaneously.