Precision limits in precision-aligned direct-drive and direct-detection architectures for photonic computing

On-chip photonic computing shows promise for tasks such as neural networks due to its parallelism and low latency. However, the impact of the nonlinear transfer curve of photonic modulators on the accuracy of photonic computing is underexplored. Current systems rely on digital post-processing, limiting energy efficiency. To address this, we propose a direct-drive and direct-detection (D4) architecture with precision-alignment design, simplifying the photonic-electronic interface. We develop a theoretical model to analyze precision constraints from modulation nonlinearity in MRRs and MZIs. Our analysis shows that both devices offer about 7 effective number of bits (ENOBs), limited by detector shot noise. This indicates that achieving high precision (e.g., 8-16 bits) for analog photonic computing faces fundamental limitations under current assumptions. This study highlights key challenges for future development.