Large-scale convolutional neural networks (CNNs) often rely on dedicated digital hardware, constrained by latency, throughput, and energy efficiency. Photonic hardware offers a promising alternative, but its analog nature and optical complexity pose challenges for electronic design automation (EDA) and design space exploration (DSE), limiting large-scale analysis. This work presents a novel parameterization methodology that quantifies the impact of noise, quantization, and kernel choice on a photonic convolution accelerator (CA), leveraging high-speed simulation tool. Using the MNIST dataset with $\mathbf{1 0}$ kernels, our results reveal up to a $3.5 \times$ difference in the root mean squared error (RMSE) between Blur and Laplacian kernels, demonstrating the critical role of kernel choice. The proposed simulation approach is also over $100 \times$ faster than conventional methods, making the analysis feasible, whereas performing it with traditional techniques would be impractical, if not impossible.

As AI continues to grow in importance, in order to reduce its carbon footprint and utilization of computer resources, numerous alternatives are under investigation to improve its hardware building blocks. In particular, in convolutional neural networks (CNNs), the convolution function represents the most important operation and one of the best targets for optimization. A new approach to convolution had recently emerged using optics, phase-change materials (PCMs) and stochastic computing, but is thus far limited to unsigned operands. In this paper, we propose an extension in which the convolutional kernels are signed, using mixed-polarity bitstreams. We present a proof of validity for our method, while also showing that, in simulation and under similar operating conditions, our approach is less affected by noise than the common approach in the literature.

State-of-the-Art (SotA) hardware implementations of Deep Neural Networks (DNNs) incur high latencies and costs. Binary Neural Networks (BNNs) are potential alternative solutions to realize faster implementations without losing accuracy. In this paper, we first present a new data mapping, called TacitMap, suited for BNNs implemented based on a Computation-In-Memory (CIM) architecture. TacitMap maximizes the use of available parallelism, while CIM architecture eliminates the data movement overhead. We then propose a hardware accelerator based on optical phase change memory (oPCM) called EinsteinBarrier. Ein-steinBarrier incorporates TacitMap and adds an extra dimension for parallelism through wavelength division multiplexing, leading to extra latency reduction. The simulation results show that, compared to the SotA CIM baseline, TacitMap and EinsteinBarrier significantly improve execution time by up to$\sim 154\times$and$\sim 3113\times$, respectively, while also maintaining the energy consumption within 60% of that in the CIM baseline.

In this paper we propose a novel approach to multiply-accumulate (MAC) operations in photonics. This approach is based on stochastic computing and on the dynamic behavior of phase-change materials (PCMs), leading to the unique characteristic of automatically storing the result in non-volatile memory. We demonstrate that, even with perfect look-up tables, the standard approach to PCM scalar multiplication is highly susceptible to perturbations as small as 0.1% of the input power, causing repetitive peaks of 600% relative error. In the same operating conditions, the proposed method achieves an average of 7× improvement in precision.

This paper investigates the potential of a compute-in-memory core based on optical Phase Change Materials (oPCMs) to speed up and reduce the energy consumption of the Matrix-Matrix-Multiplication operation. The paper also proposes a new data mapping for Binary Neural Networks (BNNs) tailored for our oPCM core. The preliminary results show a significant latency improvement irrespective of the evaluated network structure and size. The improvement varies from network to network and goes up to ~1053x.