EDA IR Drop Prediction

Update

• 2023/08/04
  Add the model overview and brief description.
• 2023/6/24
  Add the IR_drop_train.py file.
  Delete model weight
  Get model weight at GoogleDrive

Introduce

IR drop analysis is an essential step for IC design, evaluating the power integrity of a chip power delivery network. This is a deep-learning-based EDA tool for predicting the IR Drop heatmap, I achieve high precision and accuracy by inputting the power features. Additionally, for one prediction process only costs about 2 seconds.

Requirement

1. python3.8
2. scipy
3. matplotlib
4. numpy
5. opencv
6. pandas
7. pytorch 1.12.0

Model Overview

feature	discribe
$p_{i}$	Internal power
$p_{s}$	Switching power
$p_{sac}$	Toggle rate scaled power
$p_{all}$	$p_{i}+p_{s}+p_{sac}$
$p_{t}$	The power of an instance at each time step

For more feature information V. A. Chhabria et al. [1] have discribe the detail of the feature extraction process.
The CE-FPN model is proposed by Y. Luo et al. [2].

1. The purple CNN block is the CoordConv.
2. SSF is the Sub-pixel Skip Fusion.
3. SCE block is the Sub-pixel Context Enhancement.
4. CAG block is the Channel Attention Guided Module.

Train

python IR_drop_train.py

--root_path: The path of the data file
--batch_size: The batch size figure
--num_epochs: The training epochs
--learning_rate: learning rate [0,1]
--weight_path: The path to save the model weight
--fig_path: The path of the figure file

Predict

python IR_drop_predict.py

--data_path: The path of the data file
--fig_save_path: The path you want to save figure
--weight_path: The path of the model weight
--output_path: The path of the predict output with .npy file
--irdrop_threshold: irdrop_threshold [0,1]
--device: If you want to use gpu type "cuda"

Predict result

1. Tune your own irdrop_threshold, the defalt is 0.1 as shown in following figure.
2. The output coordinate csv file and image array npy file are in the ./output file.
3. The model predict cost time is 1.49 ~ 2.2 sec.

Compare with ground truth

Cross validation while evalulate the model

ROC Curve:

SSIM score: 0.863
AUC: 0.991
Percision: 0.9975
by pytorch_msssim.SSIM
by the same metrics code as CircuitNet

Reference

[1] V. A. Chhabria, Y. Zhang, H. Ren, B. Keller, and S. S. Sapatnekar, "Mavirec: mL-aided vectored ir-drop estimation and classification," *Design, Automation & Test in Europe Conference & Exhibition*, pp. 1825-1828, 2021.
[2] Y. Luo, X. Cao, J. Zhang, X. Cao, F. Guo, H. Shen, T. Wang, and Q. Feng, "Ce-fpn: enhancing channel information for object detection," ArXiv Computer Vision and Pattern Recognition, pp. 1-9, 2021.