Coherent photonic crossbar arrays for integrated artificial intelligence (AI) and machine learning (ML) accelerator system.

The dramatic recent progress in the field of artificial intelligence (AI) and machine learning (ML) is pushing persistently for unprecedented computational power which is already overwhelming for the conventional computing systems. Also, the pace of the increase in computing power requirements is growing significantly faster than the progress in the conventional computing and the Moore's law. Therefore, we need to implement novel computing techniques to meet the demands of advanced AI, ML based systems and applications.

In this scenario, computing approaches with hardware accelerators, implemented using photonic crossbar arrays and phase-change materials (PCM), can potentially accomplish low latency and high efficiency matrix operations for computationally heavy applications, such as in AI, ML and deep learning. So, an analog optical computing system based on photonic crossbar arrays and phase-change materials can be a potential candidate to achieve large compute powers (TOPS) at high energy-efficiency (TOPS/W) for AI, ML acceleration in datacenters and supercomputers

However, design and implementation of such a practical system-level architectures compatible for example, with existing datacenter, supercomputing systems is highly challenging. Issues such as scalability, on-chip integration techniques & compatibility, device footprints, power consumption etc. has to be considered and addressed carefully in the design. And corresponding fabrication, integration techniques have to be devised in order to realize a functional hardware accelerator system.

In this work, we

  1. research and identify an efficient, scalable integration route for a photonic crossbar based AI accelerator architecture.
  2. develop the optical-electrical testing setup for coherent photonic crossbar arrays
  3. design, fabricate and test a proof of principle coherent photonic crossbar (2x2) array

The work will involve extensive study of integrated photonic hardware architectures of AI/ML accelerator systems; nanophotonic component & circuit design through electromagnetic simulations and optimization. Depending on the design outcomes, it will require nanophotonic device/circuit fabrication & PCM deposition in the Kavli/EKL cleanroom for implementing the coherent photonic crossbar (2x2) array. Finally, the optical-electrical characterization of the fabricated devices, circuits will be performed depending on the progress during the project timeframe.

[1]. Nat. Phot. 9, pp. 725–732 (2015)
[2]. arXiv:2210.10851 (2022)
[3]. Nat. 589, pp. 52–58 (2021)
[4]. IEEE Journal of Selected Topics in Quantum Electronics: Optical Computing 29, 2 (2023)