Improved Crosstalk Modeling for Noise Constrained Interconnect Optimization

1
Improved Crosstalk Modeling for Noise Constrained
Interconnect Optimization

• Jason Cong, David Zhigang Pan Prasanna V.
  Srinivas
• Department of Computer Science, UCLA
• Magma Design Automation, Inc.
• 2 Results Way, Cupertino, CA 95014

2
Motivation

• Deep sub-micron net designs have higher aspect
  ratio (height /width )
• Height is not reducing not same as width is
  reducing
• Make a simple and accurate modeling to calculate
  noise analysis in circuit
• Increased coupling capacitance between nets
  compared to previous process
• Longer propagation delay or Shorter propagation
  delay
• Increased logic errors --- Glitch Analysis

3
Motivation

• Reduced noise margins
• Lower supply voltages
• Dynamic Logic
• Lower Vt
• Crosstalk cannot be ignored
• It can cause low yield and low performance

4
Aggressor / Victim Network

• Assuming idle victim net
• Ls Interconnect length before coupling
• Lc Interconnect length of coupling
• Le Interconnect length after coupling
• Aggressor has clock slew tr

5
2-p Model

• Victim net is modeled as 2 p-RC circuits
• Rd Victim drive resistance
• Cx is ased to be in middle of Lc,
• Cx is decreasing as distance is increasing

6
2-p Model Parameters
7
Analytical Solution
8
Analytical Solution part 2

• s-domain output voltage
• Transform function H(s)

9
Analytical Solution part 3

• Aggressor input signal
• Output voltage

10
Simplification of Closed Form Solution

• Closed form solution complicated
• Non-intuitive
• Noise peak amplitude, noise width?
• Dominant-pole approximation method

11
Dominant-Pole Simplification
12
Intuition of Dominant Pole Simplification

• vout rises until tr and decays after
• vmax evaluated at tr

13
Extension to RC Trees

• Similar to previous model with addition of lumped
  capacitances
• Extended to a victim net in general RC tree
  structure

14
Results

• Average errors of 4 comparing to HSPICE in peak
  noise and noise width.
• Devgan model 589
• Vittal model 9
• 95 of nets have errors less than 10

15
Spice Comparison

• peak noise noise width

16
Effect of Aggressor Location

• As aggressor is moved close to receiver, peak
  noise is increased

Ls varies from 0 to 1mm Lc has length of 1mm Le
varies from 1mm to 0
17
Optimization Rules

• Rule 1
• If RsC1 lt ReCL
• Sizing up victim driver will reduce peak noise
• If RsC1 gt ReCL and tr ltlt tv
• Driver sizing will not reduce peak noise
• Rule 2
• Noise-sensitive victims should avoid
  near-receiver coupling

18
Optimization Rules part 2

• Rule 3
• Preferred position for shield insertion is near a
  noise sensitive receiver
• Rule 4
• Wire spacing is an effective way to reduce noise
• Rule 5
• Noise amplitude-width product has lower bound
• And upper bound

19
Conclusions

• 2-p model achieves results within 6 error of
  HSPICE simulation for crosstalk noise model.
• Dominant node simplification gives intuition to
  important parameters
• Design rules established to reduce noise