Computers were born analog, with the very first implementations of AI accelerators in the 80s in the form of optical computers. Unfortunately, none of them are in use today and this is because of the exponential gains of digital computers under Moore’s law. However, this trend has now ended and, to keep up with the increasing demands due to highly important workloads, such as machine learning and hard optimization, we need alternative computing technologies.
The talk will discuss our efforts in building an optoelectronic computer called analog iterative machine (AIM), that can solve a large class of practically important optimization problems at room temperature. AIM’s design leverages commodity and chip-scale optical and electronic technologies for scalability and offers around 100 × speed-up as compared to state-of-the-art digital chips.