Stub Equalizer

1
Stub Equalizer

• Behzad Razavi

2
Outline

• Problem Statement
• Architecture I
• Programmable Delay
• Architecture II
• Combination of Linear and Stub Equalizers
• New Linear Equalizer

3
Overall View

• Spec
• Max stub length 2 inches
• Max cable attenuation 25dB at 1.7GHz
• Display port uses double side termination . Max
  data rate 6GHz
• HDMI uses single side termination . Max data rate
  3.4GHz

• Need to implement delay between
• 0ps and 600ps .

4
Impulse Response
Impulse response for stub 1.5 cm
5
Freq. Response
Freq response for stub 15 cm
6
Eye
Stub 5 mm, no equal.
7
Eye
Stub 1.3 cm, no equal.
8
Eye
Eye at 50ohm termination
The purpose is to move middle plateau region to
high plateau
Eye at summing node of page 1 .
Stub .5 cm del 0
9
Eye
Eye at summing node of page 1 .
Eye at summing node of page 1 .
This is a real transition .
Stub 15 cm del 1.84 ns
10
Eye
How to find out an error function for DFE ?
Stub 15 cm, no equal.
11
Eye
Stub 15 cm, del 1.82 ns
12
Eye
The resolution of delay cell need to be finer
than 20ps .
Stub 15 cm, del 1.86 ns
13
Simple Delay Line

• Minimum delay is about 20 ps, but we need about
  15 ps.
• Large cap at the output
• ? interpolate, but delay per stage is limited
  by required bandwidth ? still large cap at the
  output

14
Two-Step Synchronous Delay Line
15
Delay Variation
16
Min. Delay
17
Delay with Gm1 and Gm9
18
Delay with Gm1 and Gm10
19
Delay with Gm1 and Gm11
20
Delay with Gm1 and Gm13
21
Interpolation between Gm12 and Gm13
22
Delay with Gm2 and Gm9
23
Eye Opening Measurement

• Threshold must be larger than 3sigma offset.

24
Eye is open for any choice of delay!
Eye at Summing Node
Received Data
Stub 14 mm, delay 0
25
Eye at Summing Node
Received Data
Stub 15 mm, delay 0
26
Eye at Summing Node
Received Data
Stub 15 mm, delay 0.84 ns
27
Eye at Summing Node
Received Data
Stub 15 cm, delay 0.84 ns
28
Discussion

• Possible approaches
• 1. Measure BER during training and adjust
  delay accordingly.
• 2. Measure spectrum of received data and
  predict required delay.
• 3. Measure plateau of received data and
  predict required delay.

29
Plateau Measurement

• Need to find the longest sequence of ZEROs
  between two ONEs.
• Need a resolution of about 20 ps and some guard
  time ? about 15 window detectors to span 200 ps.

30
Implementation
31
Duplication of Input Signal
32
Spectrum Analyzer

• First LPF bandwidth must be small enough to
  provide a large dynamic range for the dc output.
  (40 MHz)
• Odd harmonics of LO not very serious because the
  null repeats at those harmonics.

33
5-cm Stub, LO 740 MHz
Input
Mixer Output
DC Output
34
5-cm Stub, LO760 MHz
Input
Mixer Output
DC Output
35
5-cm Stub, LO780 MHz
Input
Mixer Output
DC Output
36
5-cm Stub, LO800 MHz
Input
Mixer Output
DC Output
37
Overall Response
38
2.5-cm Stub
39
Spectrum Analyzer Requirements

• PLL covering a range of 700 MHz to 5 GHz in
  40-MHz steps (can be a ring osc.)
• Third-order LPF output energy is proportional to
  40/5000 -21 dB, e.g., -21 dBm for a 0-dBm input
• Amplify first LPF output.
• Rectifier
• ADC Storage

40
Use of Spectrum Analyzer
41
Precise Delay Setting
42
Thoughts on DFE
43
Delay 0.51 ns
XOR Output
DFE Output
44
Delay 0.49 ns
XOR Output
DFE Output
45
Delay 0.53 ns
XOR Output
DFE Output
46
Overall Behavior
47
Spectrum Analyzer Issues

• At frequencies near notches, the third harmonic
  of LO downconverts energy ? reconstructed
  spectrum becomes nonmonotonic
• The baseband dc output depends on the bit
  pattern. ? consecutive measurements may not be
  consistent.

48
Linear Equalizer Issues

• Can we put equalizer in maximum boost? No.
• Can we measure the eye opening instead of energy?
  No, the stub reflection closes the eye.
• ? Need to measure the entire spectrum and
  decide how much boost is needed.

49
Merging DFE and FFE
50
Analog Stub Equalizer

• If the equalizer does not use flipflops, then the
  output eye is closed if equalization is not
  correct.
• (Objective is to avoid the spectrum analyzer.)
• ? Approximate the inverse of the channel by a
  polynomial

51
Example

• But too noisy.

52
Passive Boost for Linear Equalizer?

• Can passive boost with maximum peaking operate
  properly?
• If the gain of amplifier is large enough to
  compensate the dc loss, it becomes nonlinear due
  to the sharp peaks.

53
Linear Equalizer Revisited

• Does the adaptation get confused?

54
Simulated System
55
Adaptive Equalizer
56
No Stub
TX Output
cont
X
57
Stub 1 cm
TX Output
cont
X
58
Stub 1.5 cm
TX Output
cont
X
59
Stub 2 cm
TX Output
cont
X
60
Stub 3.5 cm
TX Output
cont
X
61
Stub 5 cm
TX Output
cont
X
62
Analog Stub Equalizer
63
Simulated System
Linear Equalizer

• No clock!
• No need for spectrum analyzer eye closes with
• improper equalization.
• Need to check with different stub lengths and set
  D1 accordingly.
• Need to consider two extreme cases
• - Cable with maximum loss
• - No cable

64
5-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
65
5-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
66
15-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
67
25-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
68
35-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
69
45-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
70
50-mm Stub, Maximum Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
71
15-mm Stub, No Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
72
35-mm Stub, No Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
73
50-mm Stub, No Cable Loss
Received Data
Linear Equalizer Output
Stub Equalizer Output
74
Linear Equalizer Architecture

• Compensates for 30 dB loss at 2.5 GHz (50 dB
  loss
• at 5 GHz).
• Operates well even without adaptation. Optional
  path is
• added for adaptation.
• Designed in TSMC T90G.
• To monitor the output eye, give the circuit
  about 10 ns
• to reach steady state.
• Input needs 50-ohm terminations to Vdd. Input CM
  is
• 0.9 V.

75
Circuit Details (I)

• Channel lengths are 0.1 um unless otherwise
  stated.

76
Circuit Details (II)

• To adapt, change I1 and I2 such that their sum
  is 0.5 mA.

77
Channel Model

• Cascade 24 sections for a loss of 30 dB at 2.5
  GHz..

78
Output Eye at Max. Loss
79
Output Eye at Min. Loss
80
Output Eye at Mid-Loss No Aux. Path
81
Output Eye at Mid-Loss With Aux. Path
I10.4 mA I20.1 mA
82
Freq. Response
83
Equalizer Response