Scientists have designed a new microchip that's powered by light rather than electricity. The tech has the potential to train future artificial intelligence (AI) models much faster and more efficiently than today's best components, researchers claim.
By using photons to perform complex calculations, rather than electrons, the chip could overcome the limitations of classic silicon chip architecture and vastly accelerate the processing speed of computers, while also reducing their energy consumption, scientists said in a new study, published Feb. 16 in the journal Nature Photonics.
Silicon chips have transistors — or tiny electrical switches — that turn on or off when voltage is applied. Generally speaking, the more transistors a chip has, the more computing power it has — and the more power it requires to operate.
Throughout computing history, chips have adhered to Moore's Law, which states the number of transistors will double every two years without a rise in production costs or energy consumption. But there are physical limitations to silicon chips, including the maximum speed transistors can operate at, the heat they generate from resistance, and the smallest size chip scientists can make.
It means stacking billions of transistors onto increasingly small silicon-electronic chips might not be feasible as the demand for power increases in the future — particularly for power-hungry AI systems.
Using photons, however, has many advantages over electrons. Firstly, they move faster than electrons — which cannot reach the speed of light. While electrons can move at close to these speeds, such systems would need an extraordinary — and unfeasible — amount of energy. Using light would therefore be far less energy-intensive. Photons are also massless and do not emit heat in the same way that electrons carrying an electrical charge do.
In designing their chip, the scientists set out to build a light-based platform that could perform calculations known as vector-matrix multiplications. This is one of the key mathematical operations used to train neural networks — machine-learning models arranged to mimic the architecture of the human brain. AI tools like ChatGPT and Google's Gemini are trained in this way.
Instead of using a silicon wafer of uniform height for the semiconductor, as conventional silicon chips do, the scientists made the silicon thinner — but only in specific regions.
"Those variations in height — without the addition of any other materials — provide a means of controlling the propagation of light through the chip, since the variations in height can be distributed to cause light to scatter in specific patterns, allowing the chip to perform mathematical calculations at the speed of light," co-lead author Nader Engheta, professor of physics at the University of Pennsylvania, said in a statement.
The researchers claim their design can fit into pre-existing production methods without any need to adapt it. This is because the methods they used to build their photonic chip were the same as those used to make conventional chips.
They added the design schematics can be adapted for use in augmenting graphics processing units (GPUs), for which demand has skyrocketed in recent years. That's because these components are central to training large language models (LLMs) like Google's Gemini or OpenAI's ChatGPT.
"They can adopt the Silicon Photonics platform as an add-on," co-author Firooz Aflatouni, professor of electrical engineering at the University of Pennsylvania, said in the statement. "And then you could speed up [AI] training and classification."
