NEURAL NETWORK FOR EDA ROUTABILITY-DRIVEN GLOBAL ROUTING - MEDUSA Supervisor: Dr. André Ivanov Co-supervisor: Dr. Guy Lemieux Presenter: Zhonghua (Sebastian) Zhou, PhD Candidate
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
NEURAL NETWORK FOR EDA ROUTABILITY-DRIVEN GLOBAL ROUTING- MEDUSA
Supervisor: Dr. André IvanovCo-supervisor: Dr. Guy LemieuxPresenter: Zhonghua (Sebastian) Zhou, PhD Candidate
2
BACK-END PHASE
Floorplanning
Placement
Routing
Tape out
Post-layoutVerification
Floorplanning
3
A DESIGN, FROM PHYSICAL TO LOGICAL
2. Use black dots and dashed lines to represent the vertices and the edges of the grid, respectively.
3. Remove the grid, the product graph topology consists of dots and lines
1. Apply grid on top of design
1 2 3
4
ROUTING – PATH FINDING ALGORITHM
• Within a given area
• Connect source-sink pair
5
ROUTING – PATH FINDING ALGORITHM
10hrs
20mins
2hrs
6
ROUTING – PATH FINDING ALGORITHM
7
ROUTING – PATH FINDING ALGORITHM
• Routing area grows• With different grid sizes
• Source-sink pair distance grows
• Multi source-sink set
• Knowing the congestion map is of critical importance
8
PRIOR ML-BASED CONGESTION ESTIMATION
• Solves a decision problem – “whether the current tile has Design Rule Check (DRC) violation(s)”
𝑉! 𝑛 = $1, 𝑛 > 00, 𝑛 = 0
𝑛 is the total number of violations within a tile 𝑖
A Congestion Prediction Model
9
OUR WORK - MEDUSA
• MEDUSA - Machine learning congEstion preDiction USing multi-chAnnel features
• Solves a regression problem – “What’s the severity of the current tile that has routing overflows”
A Congestion Prediction Model
10
OUR WORK - FLOWCHART
11
Capacity Pin locationsNet connections Congestion
OUR WORK - FEATURES
• Image features are produced by our algorithm• Do not depend on routing tool’s outputs• Independent from input formats
12
OUR WORK - MODEL
• We aim at simplifying the machine learning model as much as possible• Advantages:
• Faster training/converging time• Lower memory usage• Faster prediction time• Easier tuning
Input Output
Encoder
13
OUR WORK - RESULTS
[2] P. Spindler and F. M. Johannes, “Fast and Accurate Routing Demand Estimation for Efficient Routability-driven Placement,” in DATE, April 2007.
[1] Z. Zhou, Z. Zhu, J. Chen, Y. Ma, B. Yu, T. Ho, G. Lemieux, and A. Ivanov, “Congestion-Aware Global Routing using Deep Convolutional Generative Adversarial Networks,” in 2019 ACM/IEEE 1st Workshop on Machine Learning for CAD (MLCAD), 2019
(a) Ground truth (b) DUDY [1] (c) Zhou [2] (d) Medusa
(a) Ground truth (b) DUDY [1] (c) Zhou [2] (d) Medusa
14
OUR WORK – RESULTS (AVG. OF 15 BENCHMARKS)
PCC (Perfect = 1.0) NRMSE (Perfect = 0.0)
RUDY Zhou Medusa RUDY Zhou Medusa
0.902 0.813 0.959 0.175 0.165 0.079
Initial Routing Overflow (x10^4) Total Routing Iterations
RUDY Zhou Medusa RUDY Zhou Medusa
Total 611.7 690.5 460.2 727 760 722
Norm. 1.329 1.500 1.000 1.006 1.053 1.000
Table I. Congestion Estimation Quality
Table II. Impact of Estimation Techniques on Routing
15
OUR WORK – UBCROUTE
• UBCRoute is a 4-stage global routing algorithm• 4 routing strategies• Different cost function focusing on different objectives
• Wirelength• Congestion• Forbidden region [3]• Post-processing
[25] H. Chen, C. Hsu, and Y. Chang, “High-performance global routingwith fast overflow reduction,” in ASP-DAC, Jan 2009
16
OUR WORK – RESULTS (AVG. OF 15 BENCHMARKS)
Runtime (s)
NCTU NTUgr2 NTHU2 UBCRoute + MEDUSA
Total 23.3e3 10.7e4 10.3e5 71.8e2
Norm. 3.252 14.859 142.967 1.000
Table III. Global Routing Runtime
17
SUMMARY
• Designed congestion estimation algorithm • Developed feature extraction method
• Independent from input format & routing tools’ output • Developed customized neural network
• Achieves high accuracy• maintains simplicity
• Impact on routability-driven routing algorithm:• More accurate congestion map• Better initial routing solution• Fewer optimization iterations• Faster global routing runtime
18
CHALLENGES
• Unknown scalability • Circuit benchmarks are relatively small• Source code of industrial routing tools are proprietary
• Integration• Most ML frameworks are Python while tools C/C++
19
ACKNOWLEDGEMENT
• UBC• HUAWEI• NSERC
• Dr. André Ivanov• Dr. Guy Lemieux• Dr. Tsung-Yi Ho (National Tsing Hua University)
Related Documents
