A light-powered computer chip could train AI much faster than semiconductors powered by electricity. The novel chip design uses photons instead of electrons to perform calculations. Scientists hope to be able to integrate this technology into future graphics cards for training AI applications.
Researchers have developed a new microchip that is powered by light rather than electricity. The technology has great potential to train future artificial intelligence (AI) models much faster and more efficiently than today's fastest components, scientists believe.
By using photons instead of electrons to perform complex calculations, the chip could overcome the limitations of classic silicon chip architecture and significantly accelerate the processing speed of computers while reducing their energy consumption, according to a group of scientists in a new study published in the journal 'Nature Photonics' in February 2024.
Silicon chips operate with transistors that are switched on and off to perform computing operations. Roughly speaking, the more transistors a chip has, the more computing power it has - and the more power it needs to operate.
Moore's law still applies
Throughout the history of computer technology, chips have always adhered to Moore's Law. It states that the number of transistors doubles every two years without production costs or energy consumption increasing. However, silicon chips are subject to physical limits. These include the maximum speed at which transistors can operate, the heat they generate through resistance and the smallest size that technologists can produce.
This means that interconnecting billions of transistors on ever-smaller silicon electronic chips may not be feasible as power requirements increase in the future - especially for energy-hungry AI systems.
The use of photons 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 this speed, such systems would require an extraordinary - and unrealizable - amount of energy. The use of light would therefore be far less energy-intensive. In addition, photons are massless and do not radiate heat, as electrons with an electric charge do.
When developing their chip, the scientists wanted to build a light-based platform that could perform calculations known as vector-matrix multiplications. This is one of the most important mathematical operations used to train neural networks - machine learning models designed to mimic the architecture of the human brain. AI tools such as ChatGPT are trained in this way.
Instead of using a silicon wafer with a uniform height for the semiconductor, as is the case with conventional silicon chips, the scientists made the silicon thinner - but only in specific areas. "These height variations - without adding other materials - provide a way to control the propagation of light through the chip because the height variations can be distributed to scatter light in specific patterns, allowing the chip to perform mathematical calculations at the speed of light," said co-author Nader Engheta, professor of physics at the University of Pennsylvania, in a brief statement.
The researchers report that their design can be integrated into existing production processes without having to be adapted. This is because the methods they used to build their photonic chip are the same as those used to make conventional chips.
They added that the design principles could also be used to augment graphics processing units (GPUs), for which demand has skyrocketed in recent years. This is because these components are central to training large language models (LLMs) such as Google Gemini or OpenAI ChatGPT.
"You can use the Silicon Photonics platform as an add-on," said co-author Firooz Aflatouni, professor of electrical engineering at the University of Pennsylvania, in the statement. "And then you could speed up the training and classification [of AI]."
www.nature.com/nphoton/
www.upenn.edu
