Short-reach Optical Communications: A Real-world Task for Neuromorphic Hardware

TL;DR

This work targets a real-world, energy-efficient benchmark for neuromorphic hardware by framing a time-dependent IM/DD optical-link task as a dataset suitable for spiking neural networks. It introduces a PyTorch-based dataset generator and two predefined parameterizations (lcdtask and ssmftask) to enable realistic evaluation of low-power, small-scale neuromorphic receivers on a practical data-center link. The approach emphasizes symbol- and bit-level demapping within PAM-4 IM/DD channels, accounting for fiber dispersion, ISI, and nonlinear photodetector effects, and compares neuromorphic receivers to established baselines under varying noise and memory requirements. By providing scalable data generation, explicit evaluation protocols, and open-source code, the work aims to accelerate energy-aware algorithm and hardware co-design for optical communications and neuromorphic processing, with a concrete path toward practical deployment.

Abstract

Spiking neural networks (SNNs) emulated on dedicated neuromorphic accelerators promise to offer energy-efficient signal processing. However, the neuromorphic advantage over traditional algorithms still remains to be demonstrated in real-world applications. Here, we describe an intensity-modulation, direct-detection (IM/DD) task that is relevant to high-speed optical communication systems used in data centers. Compared to other machine learning-inspired benchmarks, the task offers several advantages. First, the dataset is inherently time-dependent, i.e., there is a time dimension that can be natively mapped to the dynamic evolution of SNNs. Second, small-scale SNNs can achieve the target accuracy required by technical communication standards. Third, due to the small scale and the defined target accuracy, the task facilitates the optimization for real-world aspects, such as energy efficiency, resource requirements, and system complexity.

Figures (5)

• Figure 1: Sketch of a communication system.
• Figure 2: Simplified setup of an link.
• Figure 3: Exemplary transmit and receive data for the and . The left column visualizes the time-discrete transmit, respective received signal, while the right column visualizes the corresponding histograms. Both links were simulated at $\sigma_\mathrm{n}^2=-20dB$.
• Figure 4: Reference experiments for the lcdtask (A) and for the ssmftask (B). The rows in A and B show the (left column) and spiking activity (number of generated spikes, right column) over noise levels (upper row), number of hidden neurons in the network (middle row), and the number of neighboring samples $n_\textnormal{taps}$ (last row). A lower noise power results in lower . More hidden neurons show better demapping capabilities and a higher spiking activity. Because of less , the requires fewer consecutive samples $n_\textnormal{taps}$ than the . Experiments are based on arnold2023spikingvonBank2024energy.
• Figure 5: Visualization of the tensors returned by the provided datasets. A sequence of bit pairs $\bm{B}$ is mapped to symbols in $\mathcal{X}$ via indices q. The symbols are transmitted, and at the receiver $\bm{y}$ is observed. Those symbols are reshaped into overlapping chunks y_chunk containing neighboring samples to capture the information spread because of , needed to demap y[k]. The receiver infers bit decisions $\hat{b}_1[k]\hat{b}_2[k]$.