Elger Vlieg (IBM)
For the past decade, the progress in artificial intelligence (AI) has been requiring an exponential size increase of artificial neural networks (ANN), at a rate close to 10 times that of Moore’s law [1]. Consequently, the signal processing in ANNs is fueling a compute demand that is unsustainable for general-purpose computers, both in terms of energy and hardware requirements [1,2].
The analog crossbar array is an application-specific accelerator for matrix-vector multiplications (MVM), and can compute the synaptic signal transfer between fully-connected layers in ANNs [3,4]. The signal propagation between densely connected layers generates a multiply-accumulate (MAC) workload that scales quadratically with network width. Interestingly, the analog crossbar array may perform it with linear scaling on power and latency.
We have proposed a programmable integrated photonic crossbar array architecture [3], based on the photorefractive effect [5], and super-imposed refractive index gratings that act as Bragg-mirrors. This architecture is of interest, since it supports all matrix operations for ANN backpropagation training. Previously, we reported on the fabrication of photorefractive thin-film GaAs substrates, and the realization of photonic integrated circuits (PIC) therein [3]. Now, we report on the successful writing of a photorefractive synaptic connection in our prototype crossbar processor (Figure 14).
Figure 14c illustrates the analog writing of a synaptic connection between two optical signals, achieved by their interference in a photorefractive integrated optic structure. The moment that the source beam is π phase-shifted by the electro-optic phase modulator (EOPM) (purple), the diffracted signal interferes destructively with the measured destination beam (red), until the photorefractive index grating is rewritten by an exponential saturation process [3]. In conclusion, to our knowledge this work constitutes the first demonstration of the photorefractive optical effect in a PIC for neuromorphic computing. Thereby, we validated the basic operational principle of our photorefractive crossbar array design. To extend this work into a fully functional crossbar array, programmable I/O circuits need to be added to the prototype design, so that multiple I/O waveguides can addressed in parallel and individually (Figure 14a) [3]. Future work includes: the design optimization for attenuated crosstalk, the fabrication of a programmable multi-coefficient crossbar array device, and the demonstration of all the relevant signal processing.
References
[1] D. Narayanan, M. Shoeybi, J. Casper, P. LeGresley, M. Patwary, V. Korthikanti, D. Vainbrand, P. Kashinkunti, J. Bernauer, B. Catanzaro, A. Phanishayee, and M. Zaharia. ”Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM,” SC ’21 58, 1-15 (2021).
[2] E. Strubell, A. Ganesh, and A. McCallum. ”Energy and Policy Considerations for Modern Deep Learning Research,” AAAI, 34(09), 13693-13696 (2020).
[3] E. A. Vlieg, L. Talandier, R. Dangel, F. Horst, and B. J. Offrein. ”An Integrated Photorefractive Analog Matrix-Vector Multiplier for Machine Learning,” Appl. Sci. 12(09), 4226 (2022).
[4] A. Sebastian, M. Le Gallo, R. Khaddam-Aljameh, and E. Eleftheriou. ”Memory devices and applications for in-memory computing,” Nat. Nanotechnol. 15, 529–544 (2020)
[5] Y. Owechko, and B.H. Soffer. ”Holographic Neurocomputer Utilizing Laser Diode Light Source,” Proceedings of an International Conference on Optical Implementation of Information Processing, San Diego, California (United States), 10–11 Jul. 1995. pp. 12–19.