Microsoft Points to an Analog Optical Computer as a Potential Solution to AI's Energy Crisis

A team of scientists from Microsoft Research and the University of Cambridge has demonstrated a potential solution to the issue of the growing energy demanded by the artificial intelligence (AI) boom: an analog, optical computer system whose core operating principles reach back to the 1960s.

"Our breakthrough work on an analog optical computer points to new ways to solve complex real-world problems with much greater efficiency," Microsoft chair and chief executive officer Satya Nadella said of the team's work. "Super to see this published today in Nature Magazine [sic]."

The analog optical computer (AOC) built by the team is considerably more powerful than one the company announced back in 2023, handling 256 weights or parameters to the last-generation model's 64. As before, though, it's based around optical rather than electronic computing: micro-LEDs are used to fire light though lenses and fiber-optic cables to digital sensors, providing a means of working with continuous values rather than the on-or-off of a binary computer system.

Microsoft's researchers are clear about one thing, though: it's not just a theoretical proof-of-concept, but a platform that can be used to solve real-world problems in both inference for artificial intelligence tasks and combinatorial optimization problems. Tasks executed on the computer included image classification, non-linear regression, medical image reconstruction, and financial transaction settlement.

"It is an absolute giant problem with massive real-world finance impact," Hitesh Ballani, co-corresponding author and director of the Microsoft Research Lab in Cambridge, claims of this latter task, which used synthetic data for 1,800 parties across 28,000 transactions provided by Barclays Bank. "It's already a problem where banks need to collaborate, and better algorithms help everyone."

These successes must be tempered with the knowledge that, even at four times the size of 2023's AOC, the new system is still too small-scale to make a real-world impact. "To be transparent, it’s not something we can go and use clinically right now," co-author Michael Hansen, senior director of biomedical signal processing at Microsoft Health Futures, admits of the biomedical tests, which aimed to reduce the amount of MRI data required to produce a clinically useful result. "Because it's just this little small problem that we ran, but it gives you that little spark that says, 'Oh boy! If this instrument was actually in full scale…'"

The team's work has been published under open-access terms in the journal Nature.