Photon: The Solution for Ultrafast and Energy-Efficient AI?…

Photon: The Solution for Ultrafast and Energy-Efficient AI?

MIT researchers have developed an integrated “photonic” processor capable of performing all the necessary operations of a deep neural network (the basis of AI) directly on a chip. This breakthrough could pave the way for deep learning calculations and therefore for faster and more energy-efficient artificial intelligence, particularly for AI processing visual streams, but not only…

Deep neural networks, which power many modern artificial intelligence applications, have reached a size and complexity that exceed the capabilities of traditional electronic hardware. A new photonic processor, based on a decade of research, is able to overcome the limitations of previous photonic devices by integrating linear and nonlinear operations on a single chip, a first of its kind.

Photons instead of electrons
A photonic processor, sometimes called an optical processor, is a computer chip designed to process digital data using photons (carried by light signals) rather than electrons (carried by electrical signals). In such systems, data is converted into photons—the elementary particles of light—by laser diodes, then converted back into electrical signals by photodetectors for storage.

Photons—according to Albert Einstein’s special relativity—are the fastest particles in the universe. The speed of electrons depends on their energy and the medium they travel through, but in a processor is about 50% of the speed of light in a vacuum. But it’s not just speed that works in favor of photons: they also allow for increased bandwidths to carry more data at the same time.

This is why the idea of ​​using photons rather than electrons to power computers is a very active area of ​​research. For example, two years ago, Oxford University unveiled “nanowires” that respond to the polarization of light and enable the storage of optical data with extreme density. That same year, Intel unveiled its miniature optical interconnect technology. In June 2024, Intel’s Integrated Photonics Solutions group unveiled the first fully integrated coherent optical interconnect chiplet. This component enables bandwidth of 4 Tbps, while reducing power consumption and increasing the range of communications.

Replacing electrons with photons is not easy. Photons are less easily controlled. And while we already know how to manipulate linear functions (i.e. simple mathematical operations, such as matrix multiplications) with photons (the first prototypes date back to 2017, but Europe has developed such a processor called Smartlight and marketed by iPronics since mid-2024), research has so far stumbled on the use of photons for non-linear functions that are essential for modeling complex relationships and for learning AI models.

AI at the speed of light
Made up of interconnected modules forming an optical neural network, the processor created by MIT researchers can not only perform optical non-linear functions but is also designed to remain in the optical domain throughout the calculation process, up to the stage of reading the results, thus reducing energy consumption and improving speed.

In addition, it is manufactured using standard industrial processes. This should ultimately make it easier to scale up the technology and integrate it with other electronic components.

In tests, the photonic chip performed calculations in less than half a nanosecond with an accuracy of over 92%, a performance comparable to that of conventional electronic hardware. The chip could transform fields such as high-speed telecommunications, scientific research in astronomy and particle physics, and navigation systems. In addition, the chip’s ability to run optical neural networks end-to-end, with extremely low latency, opens the way to applications where computational speed is essential.

According to Saumil Bandyopadhyay, the lead author of the study published in Nature Photonics, this advance opens up new applications and algorithms, because the entire neural network now operates in optics, and at the nanosecond scale. For systems processing optical signals, such as those used for navigation or telecommunications, this photonic processor could therefore offer a solution that is both efficient and energy-efficient, which is already proving to be a necessity given the unsustainable increase in energy consumption with current AIs.

https://www.nature.com/articles/s41566-024-01567-z