The Analog Iterative Machine (AIM), developed by our collaborators at Microsoft Research Cambridge, offers a new way of computation that has the potential to surpass state-of-the-art digital technology. Featuring a fully analog design and operating at the speed of light, AIM is designed to solve difficult optimization problems.


AIM uses a sophisticated form of gradient descent that solves QUMO (quadratic unconstrained mixed optimization) problems. This can represent mixed - binary and continuous - variables, which makes it a natural choice for many real-word optimization problems. The AIM algorithm corresponds to the dynamics of a second order differential equation due to the inclusion of a momentum term from the heavy ball method. Annealing terms and nonlinearities are included on top of the standard gradient descent term, allowing the algorithm to solve difficult nonconvex optimization problems.


In the optical domain, the variables are encoded in the intensity of light sources. Each matrix element that represents the information about the input problem is encoded as the transmissivity of a single cell of an optical matrix of modulators. All of this is achieved using readily available commodity technology. With compute-in-memory operation and spatial-division multiplexed representation of variables, AIM’s design paves the path to chip-scale architecture with 100 times speed-up per unit-power over the latest GPUs for solving problems with 10,000 variables.


At the pi-computing group we are in close contact with those at Microsoft Research Cambridge working on analog optical computing. Professor Natalia Berloff was an advisor on the QUMO abstraction and algorithm used by AIM. In addition, PhD student James Cummins undertook an internship applying AIM to MRI image reconstruction. By sharing ideas and resources we are able to benchmark existing unconventional computing architectures against AIM, demonstrating the power of analog computers over classical digital hardware.


[1] Project AIM,

[2] Analog Iterative Machine (AIM): using light to solve quadratic optimization problems with mixed variables,