A new frontier awaits — computing with light

In the guts of a new type of computer, a bunch of tiny LEDs emit a green glow. Those lights have a job to do. They’re performing calculations. Right now, this math is telling the computer how to identify handwritten images of numbers.

The computer is part of a research program at Microsoft.

This device is not enclosed in a case. “You can actually put your hand inside the computer to block the light,” says Hitesh Ballani. Once he does this, he explains, “it doesn’t know what problem it’s solving.”

Suddenly, instead of naming numbers correctly, the computer spits out “random guesses,” Ballani says. Then he takes his hand out of the way. Now the streaming light once again calculates correctly. “It’s really satisfying to see,” he says.

Ballani is a computer scientist at Microsoft Research in Cambridge, England. His job involves developing this machine, an analog optical computer. As its name suggests, it uses light to compute. And being analog means its operations are the opposite of digital.

In a digital computer, signals are either 1 or 0. It’s like a light switch that can only turn on or off. An analog signal is more like a dimmer switch. Any light intensity between fully on and fully off is possible.

To be clear, light is already an important part of digital computing. Fiber-optic cables zip data back and forth between computers as beams of light. But to actually perform calculations, these computers must first convert that light into electricity.

Electronic computers are reaching their limits. It’s been getting ever harder to increase their speed and power. At the same time, new technologies such as artificial intelligence, or AI, require more and more computing power. The “demand [for this] is just exploding,” Ballani notes. These systems also devour more and more energy.

The analog optical computer and other new light-based computing technologies might help meet such demands. This tech also could help green AI and other emerging technologies by cutting how much energy they will need.

A little light math

Do you wear eyeglasses? If you do, congratulations! You have your very own optical computer.

As light from a scene hits those lenses, it morphs. The scene goes from blurry to crisp. No matter what you look at, the glasses do the math to transform the scene. “And they’re doing it at no energy cost — except the original cost of making the glass and curving it,” says Charles Roques-Carmes. He works on optical computing at Stanford University in Stanford, Calif., and at the Massachusetts Institute of Technology in Cambridge.

A lens filter also does simple math with light. A clear filter allows all light through. That’s the same as multiplying the intensity of the light by one, explains Ballani. You always get the same intensity on the other side.

A black filter blocks all light, so it’s multiplying by zero. A tinted filter, like a pair of sunglasses, blocks just some of the light. This equates to “multiplying the intensity of light by a value between 0 and 1,” he explains. You can also add with light. In a smartphone camera’s chip, the intensity of many light beams all falling on the same area combine. This brightens that spot in the image.

Multiplication or addition happen over the entire area of a filter or a camera chip. Many different light beams get altered all at once.

Doing simple operations on vast arrays of numbers just happens to be a very important part of AI. But in a digital computer’s central processing unit, or CPU, all those calculations typically happen one by one. That eats up time and energy. An optical computer can more easily multitask this type of math all at once. Note that some specialized electronic chips used in AI — called graphics processing units (GPUs) and tensor processing units (TPUs) — also can do lots of math at once.

Educators and Parents, Sign Up for The Cheat Sheet

Weekly updates to help you use Science News Explores in the learning environment

Client key* E-mail Address* Go

Thank you for signing up!

There was a problem signing you up.

Super-smart lenses

Some new optical computers take the concept of glasses or a camera lens to the next level.

Aydogan Ozcan is an engineer at the University of California, Los Angeles. His team designs sensors that he calls diffractive optical processors. “Think of them,” he says, “as futuristic lenses.”

Each processor has multiple layers of glass or some other material through which light can pass. Each layer contains many tiny structures. Each alters light differently. And although they can perform complex tasks, these processors require no energy to run. 

A factory might use one to find defects in the product it’s making.

“Let’s say I’m producing some cancer drug,” Ozcan suggests. It’s important to find any defective drugs so they don’t get delivered to patients. Almost all of the drugs will be defect-free. Yet factories today use computers to process images of every single batch, scanning for the occasional defect. That wastes time and energy.

“At the end of the day, you have too much data,” Ozcan says. “We’re buried under data.”

A smart optical processor could change that. It could trigger a camera to automatically detect defects. Then, the factory would need to photograph only the defective drugs.

Ozcan’s group designed an optical processor for this task. To test whether it might work in a real-world situation, they used square samples of silicon as a stand-in for a product that a factory might produce. Silicon is a semiconductor material found in standard computer chips. Some of these samples would have defects etched into them. The engineers had to figure out what tiny structures to add to the layers of the processor so it would find these defects.

To do this, the researchers created a virtual optical processor with random structures in its layers. They also created virtual simulations of pristine silicon squares. Then they added defects to 20,000 more virtual squares.

Next, they used machine learning on a regular computer to train their new virtual processor. They showed it the virtual samples. At first, it could only guess randomly as to whether a defect was there. But after each success or mistake, the thickness of many tiny parts of the processor (altering how much light it let through) got tweaked to make correct answers more likely in the future.

In the end, virtual light shining through defective samples emerged at a different intensity than if it passed through a good sample.

Next, the team had to test that this design would also work in the real world. They 3-D printed the processor and tested it with a set of 10 real silicon wafers. The team had already etched defects into nine of these.

The processor correctly identified all the defective wafers and ignored the normal one. Ozcan’s team reported that success in an October 2023 paper in Nature Communications.

The group has also designed optical processors for many other types of tasks. One of their latest has layers that rotate. These could help it encrypt data.

Millions of multiplications

Once a smart lens gets 3-D printed, its structures won’t change. So each new task requires designing a new lens. You may only need to reprint one or two of its many layers, Ozcan notes. But it’s not really programmable in the way a typical computer is today.

Microsoft’s analog optical computer, in contrast, can be reprogrammed. But unlike Ozcan’s smart lenses, it can’t capture light from an object or scene directly. Instead, it identifies handwritten images from digital image files, which are electronic. A component — called a modulator — converts these electronic data into light by changing the brightness of the system’s green LEDs.

Those LEDs now shine onto a chip. It’s just like the chip inside an overhead projector that a teacher might use in your classroom. This projector chip brightens or dims each ray of light in order to do the math that was called for in its programming.

“There are 4 million pixels on that chip,” notes Ballani. This means that “when light is passing through,” he says, “in theory you can do 4 million multiplications [at once].”

Those millions of multiplications happen again and again. The answer to each math problem feeds back into the system, brightening or dimming the light in new ways. When the calculation is complete, the computer converts the light back into a digital, electronic signal. And this becomes the answer to the initial problem.

So it’s a hybrid optical/digital computer.

Light looping through its LEDs and the projector chip can only perform a specific type of calculation. Other types get routed through an electronic part of the computer. This component also works very differently from the CPU in a regular computer. It processes analog electrical signals, not digital ones.

Other companies, too, are working on ways to make hybrid computers that use both digital electronics and optics.

Lightmatter, based in Mountain View, Calif., is developing a chip for AI. It incorporates lasers to do light-based multiplications. All other types of calculations would go through typical electronic chips. A Boston-based company, Lightelligence, makes a computer processor called Hummingbird. It uses light to rapidly shuttle data among different electronic components within a processor.

Teaching light new tricks

A general-purpose computer that’s wholly optical does not yet exist. And there’s a good reason. Photons of light are harder to control than the electrons that drive calculations in regular computers.

In the guts of a typical digital computer, components manage the flow of electricity. This is sort of like controlling the flow of traffic along roads. Transistors rapidly start and stop electricity or increase its rate of flow. Diodes force electricity to go in just one direction.

Most types of computations require such precise traffic control.

Photons, however, don’t obey the same rules of the road. It’s tough to make them stop and start or flow in just one direction. “Most photons don’t like to interact with their environment at all,” notes Jennifer Dionne. She’s a materials scientist at Stanford University.

Controlling photons usually requires bulky materials such as large magnets. But computer chips are super teeny. Without a way to control what light is doing at these small scales, “I think we’re going to be stuck in the dark ages of computing,” Dionne says.

Thankfully, she’s an expert in finding new ways to control light. In 2019, her team designed a new type of diode. It’s very tiny. Still, it can make light flow in just one direction. Dionne got the material of this diode to act like a magnet.

“When a photon went through the material, it felt like it was in a magnetic field,” she explains, “and hence could only go in one direction.”

Since 2019, her team has been working on materials to switch light’s intensity. “We have shown that we can switch the light on and off at very fast rates,” she says.

Her team has combined this with its new diode. Her Ph.D. student, Hamish Carr Delgado, now plans to develop this into a product. It likely wouldn’t go into an optical computer just yet. But it could be very useful in data centers, Dionne says. There, it could help direct data — moving as light — among various electronic computer components.

To keep improving computers far into the future, says Dionne, we will need a range of different approaches. Optics is clearly a promising one. But it’s far from the only one. There’s also quantum computing. And computers that mimic how brains operate.

Either of those types might include light for part of its operations. For example, the company PsiQuantum, based in Palo Alto, Calif., has developed a way to do quantum operations with single photons. They have built teeny-tiny components that produce and detect single photons. And they have made components that control photons.

So far, they’ve combined tens of thousands of these components onto silicon wafers the size of dinner plates. When you look at one of these wafers under a microscope, some parts look “kind of like alien spaceships,” says the company’s co-founder, Pete Shadbolt.

So many new approaches have been emerging. Together, they suggest how light just might transform the future of computing.