Refining TDECQ

Refining TDECQ

Piers Dawe

Mellanox

Introduction• A simple reference receiver will reduce cost in

measurement (search time for TDECQ) but also in some real receiver implementations, as explained in sun_3cd_01a_0118, which showed that more than two precursor taps is not necessary

• This presentation looks at whether 0, 1 and 2 precursors are all desirable in a reference equalizer

• Also, starts to consider how to ensure transmitter quality

P802.3cd March 2018 Refining TDECQ 2

Slides 1 to 10 are the almost the

same as dawe_022818_3cd_adhoc

Slowest time-symmetric SMF signal• A simulated signal is created with a fourth-order Bessel-

Thomson filter, bandwidth chosen to set SECQ to 3.4 dB (the highest spec limit for any SMF PMD in 802.3bs or P802.3cd)

• No noise, jitter, distortion or emphasis. Any reasonable signal must be faster than this to make room for noise, jitter and distortion. This includes 100G/s/lane signals, relative to the unit interval

• If this were a 100GBASE-LR4 signal, its TDP would be 3.2 dB: too slow (spec is 2.2 dB for 100GBASE-LR4, 2.5 dB for 100GBASE-ER4)– So we expect that real 50G/s/lane signals will be faster than this

anyway

• Most other possible responses (filter types) would have a relatively faster attack and slower decay than the time-symmetric signal– See later for discussion of chromatic and modal dispersion


•

1 1.5 2 2.5 33

3.2

3.4

3.6

3.8

4

4.2

4.4

4.6

4.8

5Estimate of SECQ (dBo)

Cursor position

Est

imate

of

SE

CQ

(d

Bo

)

Refining TDECQ

Slow, symmetrical

0 1 2 3 4 5 6-0.2

0

0.2

0.4

0.6

0.8

1

After Tx and fb/2 BT4 filter

Time (UI)

Resp

on

se

, u

nit p

uls

e

For this extremely slow signal,

positions 2 and 3 have almost the

same SECQ => don't need 2nd

precursor for this signal if we have 3rd

postcursor

-2 -1 0 1 2 3 4-1

-0.5

0

0.5

1

1.5

2Tap weights for different cursor positions

Tap position relative to cursor (UI) (direction?)

Tap

weig

ht

Same transmitter in 25G PAM2 mode, 19.34 GHz BT4

UI at 25.78125 GBd

0 0.2 0.4 0.6 0.8 1 1.2 1.4

0

0.2

0.4

0.6

0.8

1

After Tx and fb/2 BT4 filter

0 0.2 0.4 0.6 0.8 1 1.2 1.4

0

0.2

0.4

0.6

0.8

1

This waveform needs

one precursor and

three postcursors

2nd precursor would be about as strong as 3rd postcursor

BT4 filter as Tx, as slow as allowed for SMF

P802.3cd March 2018 4

Slow, nearly symmetrical

Brown red orange

TDP would be 3.2

Cursor is >1.5

0 0.2 0.4 0.6 0.8 1 1.2 1.4

-0.2

0

0.2

0.4

0.6

0.8

1

1.2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

-0.2

0

0.2

0.4

0.6

0.8

1

1.2

Eyes after reference equalizer

Cursor at position 2 Cursor at position 3

Real signals are faster than this and not so clean


Corrected histogram windows; showing thresholds

Other worst-case waveforms

• A first order filter (faster attack than decay) gave the same conclusion

• Even slower waveforms with moderate 2-tap Tx FFE – same conclusion

• Would any of these waveforms have been acceptable with the original T/2-spaced equalizer?


Chromatic dispersion?

• Could a signal be that slow AND have enough chirp on some edges (not necessarily rising vs. falling), enough to make it significantly asymmetric after the fb/2 BT4 filter? A DML?

• High chirp goes with fast edges, so such a transmitter would have a high chromatic dispersion penalty if used in a PAM2, non-equalised link

• Reasonable, or a corner case the standard and the receivers don't need to go out of their way to support?


What about MMF, with its higher TDECQ limit?

• Also, modal dispersion– Contained by the fibre and modal launch specs

– Modal bandwidth is significantly more than the reference bandwidth in the receiver

• Not addressed here – for further study


What about the opposite: fast but "dirty" signals?

• While (OMA-TDECQ) controls the net useful signal strength,

• TDECQ doesn't control the net signal quality

• Conceptually – TDECQ with Ceq fixed to a constant, would

• We need something to ensure that the small opening in the eye is a reasonable proportion of the signal size – to do the job of the VEC spec in C2M

• There is a related problem with strongly over-emphasised signals that would require "inverting" FFE settings that no copper equalizer would need

• A simple way to mitigate this problem is a minimum cursor tap weight spec, e.g. 0.9


Conclusion so far

• The reference equalizer for SMF should not include the case with two precursor taps (cursor in third position) because it would be expensive to provide in some real equalizer architectures, would add search time to TDECQ measurement, and does not benefit reasonable waveforms– There might be some super-slow waveforms

(which would have failed e.g. 100GBASE-LR4) that might get slightly worse TDECQ; marking them down will help the standard and real receivers


Other maximum-TDECQ signals

• Next, looking at the variety of bad signals and considering where the limits of compliance should be


0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Ro

ugh

nes

s an

d s

ign

al's

no

ise

p

en

alti

es (

dB

o)

Slowness penalty (dBo)

SMF TDECQ limit

No patterning ornoise

Dirty

SRS 1/2 from filtering

Construction line

SRS signals

MMF TDECQ limit

TDECQ map

Different kinds of bad signal 12

TDECQ (dBo)

Ideal waveform Half the SECQ

from filtering

Slowest, as on

slides 2 to 5

P802.3cd March 2018

<- fast slow ->

<-

open

a

fter

FF

E c

losed -

>

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Ro

ugh

nes

s an

d s

ign

al's

no

ise

p

en

alti

es (

dB

o)


SMF TDECQ limit


Dirty


Construction line

SRS signals

MMF TDECQ limit

Mismatch between SRS and real signals?


TDECQ (dBo)


from filtering

Slowest, as on

slides 2 to 5

So SRS signal must be

somewhere in this range

Where will real poor

signals be? Here?

P802.3cd March 2018

<- fast slow ->

<-

open

a

fter

FF

E c

losed -

>

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Ro

ugh

nes

s an

d s

ign

al's

no

ise

p

en

alti

es (

dB

o)


SMF TDECQ limit


Dirty


Construction line

SRS signals

MMF TDECQ limit

Don't support unrealistic bad scenarios


TDECQ (dBo)


from filtering

Slowest, as on

slides 2 to 5



A region like this

should be excluded

because it requires

strong tap weights not

useful in practice

A region like this should

be excluded because

the eye after FFE is very

closed, and small

amounts of e.g.

nonlinearity would cause

big additional penalties

(cliff edge)

Like VEC issue in C2M

<-

open

a

fter

FF

E c

losed -

>

P802.3cd March 2018

<- fast slow ->


signals be? Here?

A region like this

should be

excluded

because it

requires

significant tap

weights of the

opposite sign to

normal

"Exclusion" could be by giving signals in the red boxes

worse TDECQ scores, or by "hard" pass-fail rules

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

Ro

ugh

nes

s an

d s

ign

al's

no

ise

p

en

alti

es (

dB

o)


SMF TDECQ limit

No patterning ornoiseDirty


Construction line

SRS signals

MMF TDECQ limit

Bad ISI

Don't support unrealistic bad scenarios


TDECQ (dBo)


from filtering

Slowest, as on

slides 2 to 5



A region like this

should be excluded

because it requires

strong tap weights not

useful in practice

A region like this should

be excluded because

the eye after FFE is very

closed, and small

amounts of e.g.

nonlinearity would cause

big additional penalties

(cliff edge)

Like VEC issue in C2M

<-

open

a

fter

FF

E c

losed -

>

P802.3cd March 2018

<- fast slow ->


signals be? Here?

A region like this

should be

excluded

because it

requires

significant tap

weights of the

opposite sign to

normal

Example of a fast but

bad signal on next slide

"Exclusion" could be by giving signals in the red boxes

worse TDECQ scores, or by "hard" pass-fail rules

0 0.2 0.4 0.6 0.8 1 1.2 1.4

-0.2

0

0.2

0.4

0.6

0.8

1

1.2

A tidy-looking fast bad eye after reference equalizer

• Compare slide 5 – vertical eye opening is half as much here, although both signals have same SECQ, 3.4 dB

• Worse signals are allowed by Draft 3.1 (up and to the left on the map), even for SMF

• Worse again for MMF where the draft TDECQ limit is 4.9 dB


Like the earlier bad signal, this one needs only one precursor tap

Documents

Refining TDECQ