Researchers say we could use light-powered chips to train AI

  • With artificial intelligence and large language models only getting more popular we need to find better ways to train AI and meet its computational requirements.
  • Researchers from the University of Pennsylvania have developed a chip that uses a novel manufacturing technique that uses light to complete AI tasks.
  • The silicon photonic chip would be faster, use less energy and would be more secure than existing solutions.

Despite all the advancements over the last two years, we are still in the early days of artificial intelligence and large language models. The fact is that research and development in this sphere is hamstrung by the limitations of existing hardware.

The other problem there is that when it comes to the energy requirements for AI computing, the demand is high. Unfortunately, AI arrives as multiple nations start battling with energy problems. So, to address the energy consumption and the limitations of the traditional silicon-based chips we use today, a group of researchers have suggested a novel idea – using light to improve the speed and efficiency of AI computing.

The silicon-photonic chip is the brainchild of a group led by Professor Nader Engheta. The chip uses light to provide the compute power needed for AI.

According to Engheta in a blog published by the University of Pennsylvania’s School for Engineering and Applied Sciences, the chip in question is 150 nanometers in size but only in certain areas of the chip. This uneven and varied surface allows for light to hit different parts of the chip. This means that calculations can happen at the speed of light and one couldn’t ask for something to be faster.

Ultimately, the speed at which computation can happen means that less energy is used and with predictions that the energy use of AI is only going up, this is good news.

What’s more, is that the design of this chip is ready for commercial applications. Associate professor in Electrical and System Engineering, Firooz Aflatouni who worked on this project says that the technique can even be used in GPUs as the design was restricted by what the foundry the researchers worked with was capable of doing.

However, as the Prof explains, this chip would likely work well as a part of a wider system rather than the sole piece of silicon in a system.

“They can adopt the Silicon Photonics platform as an add-on,” says Aflatouni, “and then you could speed up training and classification.”

The silicon photonic chip also has privacy benefits. As the compute speed happens at the speed of light there would no need to store data in working memory. While Aflatouni says that this essentially makes the chip nearly impossible to hack, we know that cybercriminals are a creative bunch and much like life, they will find a way.

The findings from the research have been published in Nature Photonics this month.

We’re extremely curious to see if silicon photonic chips make it to the mainstream. With chipmakers having to work harder to eke out extra hertz from silicon, using light to improve compute, secure data and reduce energy consumption is something the wider world of computing needs, not just AI.

[Image – Brian Kostiuk on Unsplash]


About Author


Related News