New memory chip controlled by light and magnets could one day make AI computing less power-hungry

A new type of ultrafast memory uses optical signals and magnets to efficiently process and store data.

The magneto-optic memory cell design could one day reduce the energy required to power AI computing farms, researchers said. (Image credit: Brian Long, UCSB)

Researchers have developed a new type of memory cell that can both store information and do high-speed, high-efficiency calculations.

The memory cell enables users to run high-speed computations inside the memory array, researchers reported Oct. 23 in the journal Nature Photonics. The faster processing speeds and low energy consumption could help scale up data centers for artificial intelligence (AI) systems.

"There's a lot of power and a lot of energy being put into scaling up data centers or computing farms that have thousands of GPUs [graphics processing units] that are running simultaneously," study co-author Nathan Youngblood, an electrical and computer engineer at the University of Pittsburgh, told Live Science. "And the solution hasn't necessarily been to make things more efficient. It's just been to buy more and more GPUs and spend more and more power. So if optics can address some of the same problems and do it more efficiently and faster, that would hopefully result in reduced power consumption and higher throughput machine learning systems."

The new cell uses magnetic fields to direct an incoming light signal either clockwise or counterclockwise through a ring-shaped resonator, a component that intensifies light of certain wavelengths, and into one of two output ports. Depending on the intensity of light at each of the output ports, the memory cell can encode a number between zero and one, or between zero and minus one. Unlike traditional memory cells, which only encode values of zero or one in one bit of information, the new cell can encode several non-integer values, allowing it to store up to 3.5 bits per cell.


Those counterclockwise and clockwise light signals are akin to " two runners on a track that are running in opposite directions around the track, and the wind is always in the face of one and to the back of the other. One can go faster than the other," Youngblood said.. "You're comparing the speed at which those two runners are running around the track, and that allows you to basically code both positive and negative numbers."

The numbers that result from this race around the ring resonator could be used to either strengthen or weaken connections between nodes in artificial neural networks, which are machine learning algorithms that process data in ways similar to the human brain. That could help the neural network identify objects in an image, for example, Youngblood said.

Unlike traditional computers, which make calculations in a central processing unit then send results to memory, the new memory cells perform high-speed computations inside the memory array itself. In-memory computing is particularly useful for applications like artificial intelligence that need to process a lot of data very quickly, Youngblood said.

The researchers also demonstrated the endurance of the magneto-optic cells. They ran more than 2 billion write and erase cycles on the cells without observing any degradation in performance, which is a 1,000-fold improvement over past photonic memory technologies, the researchers wrote.Typical flash drives are limited to between 10,000 and 100,000 write and erase cycles, Youngblood said.

In the future, Youngblood and his colleagues hope to put multiple cells onto a computer chip and try more advanced computations.

Eventually, this technology could help mitigate the amount of power needed to run artificial intelligence systems, Youngblood said.