Artificial intelligence (AI) is changing the world, but current computer systems struggle to keep up with the huge amounts of data AI needs.
A big problem is that moving data between computer parts takes a lot of energy and slows things down.
Now, researchers at Columbia Engineering have found an exciting solution that could power the next generation of AI systems.
In a study published in Nature Photonics, Professor Keren Bergman and her team introduced a new kind of chip that combines light-based (photonic) and traditional electronic technology in a 3D structure.
This new chip can move data much faster and uses much less energy than current systems.
“We’ve created a way to transfer massive amounts of data using incredibly little energy,” said Professor Bergman.
“This solves one of the biggest challenges in making AI systems faster and more efficient.”
The breakthrough chip was developed in collaboration with Professor Alyosha Molnar’s team at Cornell University.
It includes 80 light transmitters and receivers packed into a small space, offering very high bandwidth—up to 800 gigabits per second.
Even more impressive, it only uses 120 femtojoules (a tiny amount of energy) to send each bit of data. That makes it far more energy-efficient than current chips.
What makes this chip even more special is that it’s built using parts already made in commercial factories, so it could be mass-produced at a low cost.
By combining photonic and electronic parts in 3D, the chip breaks away from the limits of flat, 2D designs and allows data to move more freely and quickly between processing units.
This means AI systems can now be designed in new ways—distributing tasks across many chips without being slowed down by energy use or delays. The new technology opens the door for faster and more powerful AI, from large models like ChatGPT to real-time systems in robots and self-driving cars.
The impact goes beyond AI. This chip could also improve supercomputers, the internet, and even how computers store and access memory. It marks the beginning of a new era of fast, energy-saving computing.
This research was a joint effort by Columbia University, Cornell University, the Air Force Research Laboratory, and Dartmouth College.
