Late Breaking Results: A Neural Network that Routes ICs

Multiterminal Pathfinding in Practical VLSI Systems with Deep Neural Networks

“…Alternatively, several important attempts to route paths with a NN exhibit limited scalability and training restrictions. For example, a generative deep learning model utilized in [15] for IC pathfinding exhibits limited maximum resolution (up to 64 × 64 pathfinding tiles) and cannot replace traditional pathfinders in practical applications. Performance of the variational autoencoder (VAE) architecture used in [15] is known to significantly decrease with an increasing input resolution.…”

“…For example, a generative deep learning model utilized in [15] for IC pathfinding exhibits limited maximum resolution (up to 64 × 64 pathfinding tiles) and cannot replace traditional pathfinders in practical applications. Performance of the variational autoencoder (VAE) architecture used in [15] is known to significantly decrease with an increasing input resolution. Furthermore, the performance of this supervised method is a strong function of the robustness and quality of the training set, yielding another primary concern.…”

“…In this work, the labeled data for the supervised training comprises a limited set of existing routed analog ICs, as hand-solved by the human experts. To obtain reasonable performance with such a limited training set, ICs has been aggressively downsampled down to 64 × 64 tiles (similar to [15]). While the model performs well in analog domain, the limitations of the training set and ML architecture make these approaches impractical for pathfinding in modern digital ICs and other applications.…”

Analog Integrated Circuit Routing Techniques: An Extensive Review

“…While previous methods extract limited features from legacy data, recent surveys suggest that M/DL may play a supporting role in routing problems only in the short term [189]. The latest studies in the digital domain [185]- [188] reveal that M/DL techniques may automatically exploit massive datasets and be trained to effectively generate the paths. These methods are attempting to learn the design styles from the dataset via CNNs or deep reinforcement learning, however, at this stage, most of the successful generative capabilities were attained on the global routing only.…”

Application of Machine Learning to Physical Design

“…The global routing problem is then converted into a standard image-to-image processing problem in [25]. To determine if the tile can be used to build the input net's routing path, a neural network is trained.…”