Artificial Intelligence Consumes Too Much Power, and New Hardware Helps it Save Energy
Innovation
Perhaps just to solve a puzzle or play a game, ARTIFICIAL intelligence requires software that runs on thousands of computers simultaneously. That's about as much energy as three nuclear power plants can generate in an hour.A team of engineers at Purdue University in the US has created a piece of hardware that learns the skills of an artificial intelligence that currently runs on a software platform. Sharing intelligent features between hardware and software could offset the energy needed to use AI in advanced applications such as driverless cars or drug development.
Technology
Shriram Ramanathan, professor of materials engineering at Purdue University, said: "Software is taking on most of the challenges in AI systems. "If you integrate intelligence into circuits outside of software, you can do things you can't do today."The development of artificial intelligence hardware is still in the early stages of research. Although researchers have demonstrated AI in some potential hardware, it has yet to meet its huge energy requirements.Ramanathan says that as AI increasingly penetrates into our daily lives, a high reliance on energy-hungry software is unsustainable. If hardware and software could share intelligence features, a piece of silicon could do more with a given amount of energy input.Ramanathan's team was the first to demonstrate artificial "tree" storage in a piece of potential hardware at room temperature. Previously, researchers have only been able to observe such storage hardware at temperatures too low for electronic devices. The hardware Ramanathan's team is developing is made up of what is known as a "quantum material". These materials are famous for having properties that classical physics cannot explain. Ramanathan's lab has been working to better understand these materials and how they can be used to solve problems in electronic devices.The software USES tree storage to distribute information among "branches," making it easier to retrieve that information as it learns a new skill or task.This approach is inspired by the way the human brain categorizes information and makes decisions.
The Value
Demonstrating these trees in materials at room temperature is an important step toward hardware that promises to ease the burden on software."This discovery opens up new areas of artificial intelligence that were previously largely ignored because it is being implemented in electronic hardware that did not exist before," says Ramanathan.The material could also help create a more natural way for humans to communicate with artificial intelligence."Protons are also our natural messengers," Zhang said. A proton-powered device could eventually be a key component for direct communication with living organisms, for example through brain implants."
