Keeping an ‘Eye’ On Better AI
The growing demand of artificial intelligence (AI) applications and cloud services has highlighted the need for more efficient computing systems to handle increasingly complex workloads than current technologies can provide.
Optical computing, also called photonic computing, is showing promise over conventional hardware by using light waves produced by lasers or other sources for data storage, data processing, or data communications.
These advantages expand the complexity of AI, but there are still limitations that researchers at the University of Pittsburgh seek to overcome. Unlike traditional computer chips, which get smaller and smaller seemingly every year, photonic computing hasn’t yet reached that ability.
“Optical processing of information holds great promise for addressing many challenges facing the field of computing,” said Sadra Rahimi Kari, a graduate student in the Youngblood Photonics Lab at Pitt and lead author. “But, integrated photonic processors are typically limited by the physical size of the processing units and the energy consumption of high-speed analog-to-digital conversion.”
A New Approach to Optical Computing
By processing temporally multiplexed optical signals using dot-product unit cells (DPUC), Youngblood’s team were able to find a novel solution.
A major advantage of these unit cells is their modularity, which allows them to be assembled together as two-dimensional arrays on a single chip. This enables efficient scaling of optical computation beyond single vector-vector dot-products to matrix-vector and even matrix-matrix multiplication.
Through multiple experiments, the unit cells successfully performed multiplication operations on both real and imaginary numbers by controlling both the phase and amplitude of the optical inputs. This method is more efficient than simply controlling the amplitude of the light and also broadens the scope of applications that can be calculated on the optical processor. For example, the researchers extended these findings to compute the covariance between stochastic bit streams, which can be used to estimate similarity between data streams in the optical domain.
“Estimating covariance between two signals in real-time is of great importance for many applications including the detection of malicious attacks in data centers,” said Rahimi Kari.
In their final experiment, the researchers were able to demonstrate a path to scaling up their platform using an off-the-shelf camera to enable general purpose matrix-matrix operations. This opens the possibility of combining the benefits of both on-chip photonics and commercial image sensors to create a photonic processor which is powerful, efficient, and economically viable.
“Our approach has the potential to enable highly efficient and scalable optical computing for a broad variety of AI applications,” said Nathan Youngblood, principal investigator and assistant professor of electrical and computer engineering at Pitt’s Swanson School of Engineering.
Other researchers on the project included:
- Nicholas A. Nobile, graduate student researcher at Pitt
- Dominique Pantin, undergraduate student researcher at Pitt
- Vivswan Shah, graduate student researcher at Pitt
The paper, “Realization of an integrated coherent photonic platform for scalable matrix operations,” was recently published in the journal Optica (DOI: 10.1364/OPTICA.507525).