SERDES Link Training with Edge Inference: Neural-Network Driven Discrete Optimization to Maximize Link Efficiency

Sunil Sudhakaran, Mert Pilanci, Mark Horowitz

Keywords: CNN, ILP, Gumbel-Softmax, discrete optimization, edge inference, affine translation, hardware, microcontroller

TL;DR: Hybrid CNN-ILP framework with a custom loss function, including Gumbel-Softmax and affine translation, to optimize quantization levels in high-speed SERDES receivers for periodic on-chip edge inference, improving high-speed link metrics

Abstract: Meeting the growing data demands of modern AI applications requires efficient, high-speed communication links. We propose an edge inference framework that dynamically optimizes non-uniform quantization levels in programmable ADC receivers. While integer linear programming (ILP) offers high-quality solutions, its significant computational cost (120 seconds per instance on high-performance CPUs) and hardware requirements make it unsuitable for on-chip use. On-chip solutions are essential for fast, periodic adjustments to track time-varying effects such as temperature drift and ensure reliable communication. To address this, we train a convolutional neural network (CNN) using ILP-generated labels, achieving a 24,000x speedup with inference on a RISC-V microcontroller. The CNN leverages a custom loss function tied to system-level metrics, reducing area metric errors from 29\% to less than 2\%. Unlike prior works embedding neural networks in the signal path, our framework adapts periodically to channel variations without disrupting communication. This enables improved error rates, energy efficiency, and a scalable pathway for on-chip edge intelligence in next-generation high-speed links.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13680