Clock Metrics

WSTS 2021 Tutorial Session

March 24, 2021
Kishan Shenoi
[email protected]

Qulsar, Inc.
90 Great Oaks Blvd,andSuite
Proprietary 204
Confidential
Slide 1 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
©San Jose, INC.
2020 QULSAR, CAAll95119
Rights Reserved.
 Synchronization Metrics (Performance)

► Mathematical Model
► Fundamental Clock Concepts and Metrics
▪ Time Error (TE) and Time Interval Error (TIE)
▪ MTIE
▪ TDEV
► Relationship between TDEV, Spectrum, and MTIE
▪ Use of TDEV to identify noise type
▪ Using TDEV for guidance on loop bandwidths

Slide 2 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Common Mathematical Models
clock (t ) = A  cos( (t ) ) = A  cos(  t + 0 +  (t ) )

signal Phase function “Clock Noise”

(radian) frequency
Mathematical time
• A: Amplitude of signal. Does not figure in timing metrics.
• 0: Initial phase. Depends on choice of time origin. Usually assumed to be 0.
• (t): Can be further decomposed into different categories such as frequency error,
frequency drift, and random noise components
• ideal periodic signal: (t) is a linear function of t ((t) ≡ 0)

1
Continuous time view x(t ) = a0 + y  t +    D  t 2 +  (t )
2 Time Error
1 Models
x(nTs ) = a0 + y  nTs +    D  (nTs ) +  (nTs )
Discrete time view 2

2
Slide 3 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Clock Metrics – Basics: Time Error
clock

xn
n=0 T
ideal

► Clock signals are (almost) periodic (nominal period ~ T)

► Time Error (Phase Error):
▪ Edges do not line up – phase error (expressed in time units)
► Time Error Sequence : {xn} or {x(n)}
▪ All clock metrics derived from time error sequence
▪ Note: the time error varies “slowly” so we can divide down to a convenient rate
(However: careful when dividing down – aliasing)
▪ Common assumption: x0 = 0.

Slide 4 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Time Interval Error
Interval of interest

(n-1) n (n+1) m
Reference (“truth”)

Clock being
analyzed
x(n) x(m)

• Consider an interval of interest (e.g. 100m dash)

• Duration measured by ideal clock (“truth”) : (m - n)∙TS
• Error in measurement of same interval by clock being
analyzed:
𝑇𝐼𝐸 𝑚, 𝑛 = 𝑥 𝑚 − 𝑥(𝑛)

Slide 5 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
Clock Metrics – MTIE and TDEV
A measure of peak-to-peak excursion expected within a given interval, t
MTIE (t is a parameter). The observation interval is scanned with a moving
window of duration t and MTIE(t) is the maximum excursion.

Given a set of N observations {x(k); k=0,1,2,…,(N-1)}, with underlying

sampling interval t0, let t = n·t0 (“window” = n samples; n = 1,2,…,N).
Peak-to-peak excursion over n samples starting with sample index i is
the worst-case TIE in this interval of n samples:
k =i + n -1 k =i + n -1
peak - to - peak (i ) = { max x(k ) - min x(k )}
k =i k =i

MTIE(n), or MTIE(t), is the largest value of this peak-to-peak excursion:

N-n k =i + n -1 k =i + n -1
MTIE (n) = max { max x(k ) - min x(k )}
i =0 k =i k =i

Slide 6 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
Clock Metrics – MTIE and TDEV

MTIE MTIE is a useful indicator of the size of buffers and for

predicting buffer overflows and underflows.
Write into buffer with clock A Read out of buffer with clock B
Buffer

Buffer size > MTIE(t) implies that overflow/underflow unlikely in any interval < t
Buffer size = MTIE(t) implies that overflow/underflow occurs approx. every t seconds
Observations regarding MTIE:
• monotonically increasing with t
• linear increase indicates freq. offset
• for small t, MTIE(t) ↔ jitter
• for medium t, MTIE(t) ↔ wander
• for large t, indicates whether “locked” (zero-slope)
t

Slide 7 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
Clock Metrics – MTIE and TDEV
TDEV A measure of stability expected over a given observation
interval, t (t is a parameter).
Given a set of N observations {x(k); k=0,1,2,…,(N-1)} with underlying Note: x(k)  xk
sampling interval t0, let t = n·t0 (“window” = n samples; n = 1,2,…,N).
2
N -3 n  n + j -1  Conventional
 x (t ) = TDEV (t ) =
1
(
   i+2n i+n i 
x - 2 x + x ) Definition
6n 2 ( N - 3n + 1) j =0  i = j 
N3 
for n =1, 2 ,...,
Second-order difference

▪ TVAR = square of TDEV N-point averaging

▪ Modified Allan Deviation (MDEV)
Sum of squares
3
 y (t ) =  x (t )
t
▪ TDEV suppresses initial phase and frequency offset and quantifies
the strength of the frequency drift and noise components {i.e. (t)}
▪ TDEV provides guidance on the noise process type

Slide 8 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Clock Metrics – MTIE and TDEV

Signal Processing Interpretation of TDEV and TVAR

{x(k)} y(k;t)
+ + TVAR
S S H(z) PWR
- - LPF
EST.  x2 (t )
Z- n Z- n

First difference representative

of average frequency over t . Average of squares
Phase offset removed.

Second difference removes frequency offset

Averaging : (low pass) filter = rectangular window of length t

Slide 9 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Noise Types, TDEV, Spectrum & MTIE
TDEV (and MDEV) for different noise types Spectrum (S(f))for different noise types

Noise Process Dependence of Dependence of Noise Process Spectrum Type

TDEV(t) on t MDEV(t) on t (power)
White PM t-(1/2) t-(3/2) White PM f0

Flicker PM t0 t-1
Flicker PM f -1
Random Walk PM = White f -2
Random Walk PM = t+(1/2) t-(1/2) FM
White FM
Flicker FM f -3
Flicker FM t+1 t0
Random Walk FM f -4
Random Walk FM t+(3/2) t+(1/2)
2
 0.75    0.3  
When linear frequency drift dominates, TDEV(t) behaves as t2 S x ( f )       x   
 f    f 
Approximate relationship between TDEV and power spectrum:
 1   0.3 
(For guidance purposes only)  x (t )     Sx  
 2.5  t   t 

f0
K  f 
 TVAR  2 0   sin 2 (nft 0 )  df
1
Approximate relationship between TDEV and MTIE: M x (t )  7  4  K1  
K1 ~ 0.75; K2 ~ 0.3 (excludes effect of transients) 0
f  f 

Slide 10 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Implication of TDEV(t) versus t

FPM
FFM and RWFM
WPM WFM

B t
A

“Phase coherence” for up to A sec. Phase Flicker Floor

 Keep PLL time constants less than A
sec.
“Frequency coherence” for up to B sec. Frequency Flicker
 Keep FLL time constants less than B Floor
sec.
Slide 11 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
 Synchronization Metrics (Performance)
Topics Covered
► Mathematical Model
► Fundamental Clock Concepts and Metrics
▪ Time Error (TE) and Time Interval Error (TIE)
▪ MTIE
▪ TDEV
► Relationship between TDEV, Spectrum, and MTIE
▪ Use of TDEV to identify noise type
▪ Using TDEV for guidance on loop bandwidths

Slide 12 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.
Thank You
Questions, comments, suggestions?
[email protected]

Slide 13 Not for distribution. © 2020 QULSAR, INC. All Rights Reserved.