Microsoft Computer Can Run AI At 100x Faster

Microsoft Research has built a prototype computer that doesn’t rely on electrons zipping through silicon but on beams of light. The machine, called an analog optical computer (AOC), is designed to solve complex optimization problems and could one day handle artificial intelligence (AI) workloads with far greater speed and efficiency than today’s processors.

Unlike digital computers that crunch information in binary, the AOC embodies computations in physical systems. This avoids bottlenecks that slow down conventional chips and could make the system 100 times faster and more energy-efficient at specific tasks.

The prototype is built using commercially available parts, including micro-LED lights, optical lenses, and sensors from smartphone cameras, keeping costs down and making future mass production feasible.

The team also developed a "digital twin," a software version of the AOC that mimics the way the hardware behaves.

This allows researchers to test the system at scale, explore how optimization or AI problems would map onto the hardware, and share results with outside collaborators.

"To have the kind of success we are dreaming about, we need other researchers to be experimenting and thinking about how this hardware can be used," said Francesca Parmigiani, who leads the project at Microsoft Research Cambridge.

One of the first demonstrations involved finance. Microsoft worked with Barclays Bank to test how the AOC could optimize delivery-versus-payment securities transactions, a process used by clearinghouses to settle trades between banks.

The experimental setup handled thousands of transactions among up to 1,800 parties, just a fraction of the scale in real clearinghouses but enough to show how future versions of the hardware could make a difference.

"It is an absolute giant problem with massive real-world finance impact," said Hitesh Ballani, who directs research on future AI infrastructure at Microsoft Research.

Healthcare offered another proof-point. Using the AOC’s digital twin, researchers reconstructed MRI scans with promising accuracy. Today, a typical scan takes about 30 minutes; the optical system could theoretically cut that to just five.

While not yet ready for clinical use, the experiment hints at how the technology might reduce waiting times and improve access to diagnostics.

Michael Hansen, a senior director at Microsoft Health Futures, noted that the digital twin was crucial in showing the viability of larger-scale MRI reconstructions.