Abstract
Photonic computing has emerged as a promising solution for accelerating computation-intensive artificial intelligence (AI) workloads, offering unparalleled speed and energy efficiency, especially in resource-limited, latency-sensitive edge computing environments. However, the deployment of analog photonic tensor accelerators encounters reliability challenges due to hardware noise and environmental variations. While off-chip noise-aware training and on-chip training have been proposed to enhance the variation tolerance of optical neural accelerators with moderate, static noise, we observe a notable performance degradation over time due to temporally drifting variations, which requires a real-time, in-situ calibration mechanism. To tackle this challenging reliability issues, for the first time, we propose a lightweight dynamic on-chip remediation framework, dubbed DOCTOR, providing adaptive, in-situ accuracy recovery against temporally drifting noise. The DOCTOR framework intelligently monitors the chip status using adaptive probing and performs fast in-situ training-free calibration to restore accuracy when necessary. Recognizing nonuniform spatial variation distributions across devices and tensor cores, we also propose a variation-aware architectural remapping strategy to avoid executing critical tasks on noisy devices. Extensive experiments show that our proposed framework can guarantee sustained performance under drifting variations with 34% higher accuracy and 2-3 orders-of-magnitude lower overhead compared to state-of-the-art on-chip training methods. Our code is open-sourced at https://github.com/ScopeX-ASU/DOCTOR.
The Challenge of Temporal Drift
Analog photonic computing uses light instead of electricity to perform the massive matrix multiplications required by AI. While this is incredibly fast and energy-efficient, it is also notoriously fragile. The core components—Micro-Ring Resonators (MRRs)—are extremely sensitive to their environment. Even a tiny temperature shift (as small as 0.05K) or electrical crosstalk can cause the "weights" stored on the chip to drift, leading to a total collapse in AI model accuracy.
Existing solutions either try to "toughen up" models during initial training (which fails when noise changes over time) or use on-chip retraining. However, retraining is often too slow, power-hungry, and requires labeled data. There is a critical gap: the need for a real-time, data-free, and lightweight way to fix these drifting errors while the chip is actually in use.
The DOCTOR Framework
The authors introduce the DOCTOR (Dynamic On-Chip Temporal Variation Remediation) framework, which provides adaptive, in-situ accuracy recovery. The framework consists of three main pillars:
- Salience-Aware Sparse Calibration: Instead of retraining the entire neural network, DOCTOR treats calibration as a local regression problem. It identifies which weight blocks are most "salient" (important to the final output) and prioritizes fixing those first. Crucially, this is data-free: it probes the hardware by sending identity matrices through the chip to see how the output is distorted, then calculates a quick correction.
- Variation-Aware Tile Remapping: Recognizing that noise isn't spread evenly across a chip (e.g., one corner might be hotter than another), DOCTOR solves a "Linear Assignment Problem." It dynamically moves the most important parts of the AI model to the "cleanest" (least noisy) physical cores on the chip.
- Adaptive Remediation Controller: To avoid wasting energy, a controller periodically monitors temperature sensors and error rates. It only triggers a "remediation" (a fix) when the drift exceeds a specific threshold, balancing the trade-off between accuracy and system throughput.
Experimental Results and Efficiency
The framework was evaluated using standard computer vision benchmarks (Fashion-MNIST, CIFAR-10, CIFAR-100) mapped onto simulated photonic hardware. The authors simulated various "drifting" scenarios, including linear temperature increases and "corner hotspots."
Key results include:
- Accuracy Recovery: In scenarios where drifting noise crashed model accuracy from ~91% to ~57%, DOCTOR restored it to over 90% in just a few steps.
- Superiority over SOTA: Compared to state-of-the-art on-chip training methods, DOCTOR provided 34% higher accuracy on average.
- Massive Efficiency Gains: DOCTOR is 2 to 3 orders of magnitude (100x–1000x) faster and less resource-intensive than retraining-based methods.
- Low Overhead: The framework only adds between 0.14% and 5% to the total cycle count of the inference process, making it viable for real-world edge deployment.
Conclusion
DOCTOR represents a significant step toward reliable, high-performance photonic AI accelerators. By decoupling calibration from training data and exploiting spatial noise intelligence, it achieves robust performance with minimal overhead. While currently optimized for MRR-based architectures, the principles of salience-aware calibration and variation-aware remapping offer a blueprint for future self-correcting analog hardware.
Code available at: https://github.com/ScopeX-ASU/DOCTOR
Summary updated