News
Microsoft Tests Analog Optical Computer on Real-World Problems
Microsoft has built a prototype computer that uses light and analog signals instead of traditional binary computing. The system, developed in Cambridge, U.K., is designed to handle optimization tasks more efficiently, with potential gains in speed and energy use.
In a recent demonstration, the AOC tackled two domain-specific problems: interbank transaction settlement and MRI scan reconstruction. The system, which uses micro-LEDs, optical lenses and CMOS sensors, offloaded parts of the computation to a software "digital twin," allowing the team to simulate results at scales beyond the prototype's current physical capacity.
The financial use case was conducted in collaboration with Barclays and modeled a delivery-versus-payment (DvP) algorithm for settling tens of thousands of transactions between 1,800 parties. According to Microsoft, the results marked the first time an analog optical computer has been used to solve a real problem.
Francesca Parmigiani, senior researcher at Microsoft, noted the significance of these initial trials. "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," she said.
In the healthcare example, the team applied the system to reconstruct MRI images using sparse data, simulating a potential reduction in scan time from 30 minutes to about five. These experiments were carried out on the digital twin, which mimics the hardware's behavior to enable larger-scale testing.
Researchers also evaluated whether the same architecture could handle machine learning tasks. In a follow-up experiment, the team demonstrated that the AOC could represent memory-like states across multiple solution steps -- an important component of reasoning tasks in AI.
The system is part of Project AIM (Analog Iterative Machine), Microsoft's broader initiative to co-develop optical hardware and software for future computing. The team's optimization algorithm, a variation of a quadratic unconstrained mixed optimization (QUMO) solver, and the digital twin have been made publicly available.
While still in the early stages, Microsoft believes AOC systems like this one could eventually support AI inference workloads or domain-specific optimization where energy efficiency and speed are critical. Hitesh Ballani, who leads the effort at Microsoft, called the work a "game changer" that demonstrates the feasibility of optical analog systems for real-world use.