Microsoft has teamed up with Barclays on a novel approach to tackling artificial intelligence (AI) and optimisation problems based on a scalable analogue optical computer (AOC) architecture, designed to use consumer-grade technology.
The approach, described in a paper published in Nature, overcomes the Von Neumann bottleneck that occurs in classical computing architecture, where performance is limited by the speed with which data can move between memory and the central processing unit. The authors of the paper discuss an approach that eliminates digital-analogue conversions and the need to merge compute and memory.
In doing so, they said the AOC can achieve substantial efficiency gains in certain circumstances. They projected that the AOC would deliver performance of around 500 tera-operations per second per watt at 8-bit precision, which they claim would make it over 100 times more efficient than leading graphics processing units.
The Microsoft researchers combined 3D optical and analogue electronic technologies, using projectors with optical lenses, digital sensors and micro light-emitting diodes to build it. As the light passes through the sensor at different intensities, the AOC can add and multiply numbers, which is the basis for solving optimisation problems.
In the Nature paper, the researchers describe how the AOC hardware is able to accelerate all the compute operations to perform five key operations required for solving optimisation problems: matrix-vector multiplication, nonlinearity, annealing, addition and subtraction.
The paper also discusses how Microsoft worked with Barclays to try to solve the delivery-versus-payment securities problem, which aims to find the most efficient way to settle financial obligations between multiple parties in compliance with regulations while minimising costs or risks within the constraints of time and the balances available.
Shrirang Khedekar, a senior software engineer with the advanced technologies department at Barclays, worked with the Microsoft UK Research team to create the dataset and parameters used in the research. Khedekar, who co-authored the Nature paper, said Barclays is interested in continuing to solve optimisation problems as the capacity of future generations of the AOC grows. “We believe there is a significant potential to explore,” he said. “We have other optimisation problems as well in the financial industry, and we believe that AOC technology could potentially play a role in solving these.”
Microsoft said it would be publicly sharing its “optimisation solver” algorithm and the digital twin it developed to enable researchers from other organisations to investigate its approach to analogue optical computing.
Francesca Parmigiani, a Microsoft principal research manager who leads the team developing the AOC, said that while it is “not a general purpose computer, what we believe is that we can find a wide range of applications and real-world problems where the computer can be extremely successful”.
She said the digital twin mimics how the real AOC behaves by simulating the same inputs, processes and outputs of an AOC in a digital environment.
“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 Parmigiani.
