Congestion Estimation During TopDown Placement

1
Congestion Estimation During Top-Down Placement

• Xiaojian Yang Ryan Kastner Majid
  Sarrafzadeh
• Embedded and Reconfigurable System Lab
• Computer Science Department, UCLA

2
Outline

• Introduction
• Motivation
• Peak Congestion Prediction
• Regional Congestion Estimation
• Experimental Results
• Conclusion

3
Introduction

• Place Route Objectives
• Routability and Timing
• Placement
• Minimizing Bounding Box Wirelength
• Shorter Bounding Box ? Better Routability
• Congestion
• Routability problem
• Detours --- Timing problem

4
Motivation of Congestion Est.

• Early stages of Top-down Placement
• Logic design
• Congestion Relieving in Top-down Placement

5
Motivation of Congestion Est.

• Congestion Relieving based on estimation
• White space re-allocation
• Moving cells out of congested area

6
Basis of Estimation
Rents Rule
P T B r
P - Number of external terminals B Number of
cells T Rent coefficient r Rent exponent
7
Peak Congestion Estimation --- Worst Case
8
Peak Congestion Estimation --- Uniform
Distribution
C1
9
Peak Congestion Estimation Result
10
Peak Congestion Estimation Result
11
Regional Congestion Est.
Internal routing demand
External routing demand
Uniformly distributed routing supply
12
Internal Routing Estimation

• Wirelength Estimation based on Rents rule
• P TB
• Rent exponent r
• Locality of Rents rule
• Different subcircuits have different Rent
  Exponents
• Rent Exponent Extraction
• Dynamic extraction using partitioning tool
• Linear regression on data points
• Wirelength Estimation Model
• Donaths (1979) and Daviss (1998)

r
13
External Routing Estimation
1.0
Routing demand caused by inter-block connection
Probability-matrix within the Bounding box
14
Regional Congestion Est.
External Routing demand (routing estimation)

Internal Routing demand (wirelength estimation)

Routing demand (congestion) Of a region
15
Region Congestion Est. Experiments
Top-down Placement
64 x 64 or 128 x 128
16
Estimation Result

• 8 benchmarks, 12k cells --- 147k cells
• 2 x 2 regions
• Wirelength Estimation only 9
• Including External Routing demand 8
• 4 x 4 regions
• Wirelength Estimation only 13
• Including External Routing demand 9
• Running time
• Partitioning speed
• 147k cells, 2 x 2, 860 seconds, Sun Ultra-10
• Place / Route 8000 seconds

17
Conclusion Future Work

• Possibility to estimate congestion by Rents rule
• Congestion can be estimated during Top-down
  placement
• Peak congestion after L-shape routing can be
  accurately estimated
• Regional congestion estimation is within 10
  comparing with actual congestion by place/route
• Future work
• More accurate model for hot spot estimation
• Fast estimation by Rent parameter prediction