Re: importance of DC balance

[Jaime Kardontchik <kardontchik.jaime@ulinear.com>]

petarp@us.ibm.com wrote:

> Jaime,
>
> The noise power and not the noise rms amplitude is proportional
> to the bandwidth. Therefore, the noise due to the required extra
> bandwidth would reduce the SNR by 4 dB.

Peter,

I completely agree with you. It is exactly what I said in my
presentation in Idaho (see page 9, titled "Optical SNR")

Add now another 4 dB loss in signal power, compared with
1000BASE-X, because the bit energy is 4 dB smaller, and
you get the total 8 dB penalty in SNR of 8b/10b + 4-WDM
compared to 1000BASE-X.

I use the SNR of 1000BASE-X as a good reference to compare
different proposals for 10-Gigabit Ethernet (see my presentation
in Idaho).

Using the same reference, the 10G-BASE-T approach without
any coding gain has a 6 dB penalty in SNR. Using EVEN coding
the penalty is only 3 dB and using Trellis Coding the penalty
is 0 dB.

I summarized this in Idaho in a Table that I reproduce here
again (page 20):

                THE COMPLETE SNR Table

    Architecture                                            SNR [1]

    1000BASE-X                                         - 1 dB [2]

    8b/10b + 4-WDM                                  - 9 dB

    10G-BASE-T Even coding                   - 4 dB

    10G-BASE-T Trellis coding                - 1 dB

[1] relative units
[2] 10*log(1*0.8) ~ - 1 dB


petarp@us.ibm.com wrote:

> Jaime,
>
> I would also like to point out that in making your argument
> against 8B/10B you are comparing apples to oranges. ...
> The more meaningful comparison would be to compare
> with ...  the uncoded 2.5 Gb/s or 10Gb/s"

Peter,

On the contrary: I stressed in my presentation that I want to
compare apples to apples. Apples to apples means: take two
complete solutions that deliver 10 Gbps data. One solution
(1000BASE-T PCS + 4-WDM) uses the bandwidth very
efficiently. The other (8b/10b + 4-WDM) not. And you
see this efficiency when you compare their SNRs.

The  "uncoded (2.5 Gb/s or 10Gb/s)" solution that you refer
to is not a solution for 10 Gbps Ethernet. Without adding any
coding it cannot be used. It could lead, for example, to infinite
running lengths of 'ones'. Therefore, comparing anything
to it would be meaningless.

Jaime

Jaime E. Kardontchik
Micro Linear
San Jose, CA 95131
email: kardontchik.jaime@ulinear.com


> On June 8th Jaime Kardontchik wrote:
>
> Looking at the system from another point of
> view: static SNR. Even neglecting any high
> frequency effects on the electronics and optics,
> 8b/10b+4-WDM incurres a penalty of 8 dB in
> SNR compared to 1000BASE-X just because the
> bit energy is smaller (smaller baud period,
> 0.32 nsec vs 0.8 nsec) and the thermal noise
> is larger (3.125 GHz bandwidth vs 1.25 GHz
> bandwidth of the receiver analog front end).
> This increases either the BER and/or reduces
> the achievable link length of the 8b/10 + 4-WDM
> system.
>
> Jaime,
>
> The noise power and not the noise rms amplitude is proportional to the
> bandwidth. Therefore, the noise due to the required extra banwidth would reduce
> the SNR by 4 dB. I would also like to point out that in making your argument
> against 8B/10B you are comparing apples to oranges. You are comparing systems
> with two different line speeds, 1.25 vs. 3.125 Gb/s. The more meaningful
> comparison would be to compare two systems with the same information carrying
> capacity, for example one 8B/10B coded (let say 3.125 or 12.5 Gb/s) and the
> other uncoded (2.5 Gb/s or 10Gb/s). Therefore, the extra thermal noise would
> reduce the SNR by 1 dB.
>
> Petar Pepeljugoski
> IBM Research
> P.O.Box 218
> Yorktown Heights, NY 10598
>
> e-mail: petarp@watson.ibm.com
> phone: (914)-945-3761
> fax:        (914)-945-4134
>
> Jaime Kardontchik <kardontchik.jaime@ulinear.com> on 06/08/99 07:55:59 PM
>
> To:   stds-802-3-hssg@ieee.org
> cc:    (bcc: Petar Pepeljugoski/Watson/IBM)
> Subject:  importance of DC balance
>
> subject: Importance of DC Balance
>
> On June 4, Albert X. Widmer, from IBM, wrote:
>
> > How important is DC balance ? This question is
> > best answered by the engineers who design the
> > critical three circuits (Laser Driver, Receiver
> > Preamplifier, Clock Recovery), the persons who
> > package the electrical and optical components,
> > and those who design the verification and
> > production tests. Given an option, they generally
> > prefer a code with DC balance and a short run
> > length. After consultation with colleagues
> > active in those endeavors, I can offer the
> > following list of circuit related advantages
> > of a transmission code such as the Fiber Channel
> > 8B/10B code:
>
> > ... (a list of important technical arguments)
>
> Albert,
>
> I think that the engineers were not given
> the right options nor were asked the right
> questions. Restricting the issue to "what coding
> would you prefer from the point of view of
> DC balance ?" will give you the wrong answer
> to the main issues. (Personally, my answer
> would be: Manchester encoding. It has one
> transition per baud and the clock is already
> embedded in the data...).
>
> The answers could be very different if the
> same engineers are asked questions of the
> following type:
>
> "Give me system architecture alternatives that
> can deliver 10 Gbps with a BER of 10^(-12) in a
> LAN environment with a cost of about 3 times
> the cost of Gigabit Ethernet over fiber and
> using the installed fiber for Gigabit Ethernet".
>
> (system objectives ...)
>
> I think that we have to recognize that
>
>  "Bandwidth is not free anymore in optical
>   fiber when we go to 10 Gbps"
>
> Installed optical fiber for Gigabit Ethernet
> begins to look in some respects as the
> telephone wire when we want to increase the
> data rate by a factor of 10. Dispersion in
> the fiber becomes important. See, for example,
> Paul Kolesar's presentation in Idaho, showing
> in slide # 15 the catastrophic effects of modal
> dispersion in a traditional MMF fiber if used
> in a system at 12.5 Gbaud/sec (8b/10b + serial)
> or even in a system at 3.125 Gbaud/sec
> (8b/10b + 4-WDM, 0.32 nsec baud period): the
> modal dispersion in the fiber eats all the
> timing budget.
>
> Looking at the system from another point of
> view: static SNR. Even neglecting any high
> frequency effects on the electronics and optics,
> 8b/10b+4-WDM incurres a penalty of 8 dB in
> SNR compared to 1000BASE-X just because the
> bit energy is smaller (smaller baud period,
> 0.32 nsec vs 0.8 nsec) and the thermal noise
> is larger (3.125 GHz bandwidth vs 1.25 GHz
> bandwidth of the receiver analog front end).
> This increases either the BER and/or reduces
> the achievable link length of the 8b/10 + 4-WDM
> system.
>
> May be that at 10 Gbps we might begin looking
> at the type of techniques used to increase
> the data rate over telephone wires: multi-level
> signalling and coding gain in order to
> restrict the transmitted signal to smaller
> bandwidths.
>
> We have to judge proposals by defining first
> the overall objectives and then looking at the
> complete architecture of every proposal and
> how it balances the requirements on the
> electronics, optics and transmission medium.
>
> I do not say that 8b/10b encoding does not
> deserve a place in 10 Gbps Ethernet. We might
> define different objectives/options within the
> 10 Gbps standard (as we already have options
> within the present 1000-BASE-X Ethernet standard),
> and for some of these options 8b/10b might be
> the right choice.
>
> Jaime
>
> Jaime E. Kardontchik
> Micro Linear
> San Jose, CA 95131
> email: kardontchik.jaime@ulinear.com