NeuralSteiner: Learning Steiner Tree for Overflow-avoiding Global Routing in Chip Design

Authors: RUIZHI LIU, ZhishengZeng , Shizhe Ding, Jingyan Sui, Xingquan Li, Dongbo Bu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments over public large-scale benchmarks reveal that, compared with the state-of-the-art deep generative methods, Neural Steiner achieves up to a 99.8% reduction in overflow while speeding up the generation and maintaining a slight wirelength loss within only 1.8%.
Researcher Affiliation Academia 1SKLP, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China {liuruizhi19s, dingshizhe19s, suijingyan18b, dbu}@ict.ac.cn 2University of Chinese Academy of Sciences, Beijing 101408, China 3Central China Artificial Intelligence Research Institute, Henan Academy of Sciences, Zhengzhou 450046, Henan, China 4Beijing Institute of Open Source Chip, Beijing 100089, China 5Peng Cheng Laboratory, Shenzhen 518000, Guangdong, China {zengzhsh, lixq01}@pcl.ac.cn 6School of Mathematics and Statistics, Minnan Normal University, Zhangzhou 363000, Fujian, China
Pseudocode Yes Algorithm 1 Training in candidate point prediction phase; Algorithm 2 Parallel routing tasks construction; Algorithm 3 Net Augmented Graph Construction; Algorithm 4 Overflow-Avoiding RST Construction; Algorithm 5 Inference Algorithm
Open Source Code Yes Our source code and dataset will be released at https://github.com/ liuruizhi96/Neural Steiner
Open Datasets Yes For training, we construct the training set from ISPD07[30] using the method described in Sec. 3.3. ... We utilize a state-of-the-art traditional global router named CUGR [24] to perform routing on public benchmarks [30] also used by [8] and extract the overflow map and pin map of every net in real-time.
Dataset Splits No The paper mentions using "validation loss" to adjust the learning rate and stop training (e.g., "reduce the learning rates by 0.5 if the validation loss does not decrease in 2 epochs"), but it does not specify the method or ratio for splitting the data into a validation set.
Hardware Specification Yes Each experiment in this work is conducted on a system equipped with an Intel(R) Xeon(R) Gold 6230R CPU, NVIDIA A800 (80 GB) GPU, and 250 GB RAM.
Software Dependencies No The paper mentions software components such as CUGR, ResNet, and refers to machine learning frameworks implicitly (e.g., through terms like "neural network", "convolutional neural network"), but it does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We use learning rate in [0.001, 0.00001] and reduce the learning rates by 0.5 if the validation loss does not decrease in 2 epochs. The training will continue until the validation loss no longer decreases for over 10 epochs or the number of epoch reaches a maximum of 100. In the training loss, we use cfl = 1.0, cdi = 1.0 and cof = 2.0 to encourage the exploration of lower-overflow candidate points. The number of Res Net blocks is 8 before the RCCA module and 8 behind it.