RE: PAM-5 at 5 Gbaud
Hi Jaime and Wenbin Jiang:
There are many issues remained to be resolved; namely, optical issues, data
recovery issues, and system bandwidth issues for PAM-5 approach.
I agree with Jaime, we should concentrate in a basic model first to develop
technologies to assure that PAM-5 is feasible and reliable, before July
decision date. Without a solid demonstration of the feasibility, both
analytically and experimentally, the new approach to implement a multilevel
coding scheme will not get a wide market acceptance. Market needs a
cost-effective and reliable solution.
The optical issues have been widely discussed; therefore, it is not the
major concern.
The other issues:
1. Bit level and timing detecting.
As it was mentioned before, to recover the bit timing and level information
simultaneously from the 1.25 G baud receiving signals to meet 10^-12 BER is
not a easy job, which has not been proved in the field yet. To assume a
fiber will automatically achieve 10^-12 is a desirable objective, but not an
assurance. For example, the laser non-linearity, slope efficiency
variation, or chirping, cannot minimized or removed easily. A filter to
minimize those problems will need a RC time-constant, far larger than 266 ps
(800/3, due to three transitions in a T = 1/1.2 Gbps data). This large RC
constant filter will affect the accuracy of the ADC to detect the voltage
level with its associated timing information, because, the input signal to
the ADC should have rise time of less than 167 ps. These two requirements
are exclusive each other. It implies the laser power variation problem
cannot be resolved by a simple filtering technique. We cannot count on the
receiver TIA filter to remove it, because the receiver BW will be near 3
GHz. We need a more complex method. What is it?
2. System Bandwidth
There are three maximum transitions in a period (1/Baud rate). Therefore,
each transition time is 266 ps = 800/3 ps. Any transition time exceeds this
limit will cause additional timing distortion over the specified total TJ.
Therefore, 266 ps is the equivalent bit-cell time of a NRZ encoding.
We can apply the system bandwidth equation:
0.8x (Cell time) = rms Sum of (optical rise time + fiber rise time +
receiver rise time)
Or
0.8x266 = > (tT^2 + tF^2 + tR^2)^0.5 ..................................(1)
For a 100 meter MM fiber at 160 MHz-km, the fiber rise time, tF:
tF = (440x100) / (160x1000) = 275 ps ..................................(2)
From (1) - (2)
Obtain (tT^2 + tR^2)^0.5 = -174
...........................................(3)
The minus sign indicates that the solution is not a real number. The cable
has to be shorter to satisfy the system bandwidth equation.
If we violate the equation (1), we will get additional errors depending on
data patters and the actual bandwidth of each component. It implies the
actual link BER is larger than 10^-12.
A higher bandwidth fiber can be used to resolve the problem. Otherwise, a
high pass filter can be added to extend the bandwidth; however, this is not
just to equalize the timing distortion caused by ISI, it also needs the
accurate amplitude equalization due to the multilevel detection requirement.
We need a solution to demonstrate the reliable solution.
We should continue to resolve the PAM-5 issues from all directions.
Regards,
Edward S. Chang
NetWorth Technologies, Inc.
EChang@xxxxxxxxxxxxxxxx
Tel: (610)292-2870
Fax: (610)292-2872
-----Original Message-----
Wenbin Jiang wrote:
> Laser linearity is one concern. What about its slope efficiency variation
> with the temperature?
>
Hello Wenbin,
Under "laser linearity" I implicitly included all the laser-related
issues, including the slope efficiency variation.
The present diversity of PAM-5 proposals precludes a look
in depth at this issue: in the last Dallas meeting, if you look
at the spreadsheet, we had 1.25, 2.5 and 5 Gbaud rates and
850 nm and 1300 nm lasers.
If we will reduce this multiplictiy into one and only one
PAM-5 proposal in March, then we will be able to take a
detailed look at the laser-related issues during the next
four months and come to the July 2000 meeting with a good
answer - one way or the other.
Jaime
Jaime E. Kardontchik
Micro Linear
San Jose, CA 95131
Laser linearity is one concern. What about its slope efficiency variation
>with the temperature?
>
>Wenbin Jiang
>E2O Communications, Inc.
>26679 W. Agoura Road
>Calabasas, CA 91302
>
............................................................................
Dear Kevin:
Thanks for your interest in my comment.
Yes, I agree with you, there is a deskew issue that is not quite explored
yet. However, deskewing is a well practiced technique, and even with a
dozen bits of FIFO depth is not expensive. Especially, if it becomes a
standard part, the price will be in several dollars range, I hope.
Regards,
Edward S. Chang
NetWorth Technologies, Inc.
EChang@xxxxxxxxxxxxxxxx
Tel: (610)292-2870
Fax: (610)292-2872
-----Original Message-----
From: kdemsky@xxxxxxxxxx [mailto:kdemsky@xxxxxxxxxx]
Sent: Wednesday, February 23, 2000 1:31 PM
To: Edward Chang
Subject: RE: PAM-5 at 5 Gbaud
Mr. Chang,
I am a Fiber Channel representative who is very interested in what you guys
are doing, but not a member of 802.3, so forgive me if I dont post to
reflector.
Your comments are as valuable, and well thought out as Jaime's PAM-5
proposal discussion. I have one point of light to add to your comments
though.
You said,
"There is another way to achieve an 850 nm, 150 meter installed MM fiber
solution without those issues mentioned above by using 8B/10B-4WDM method.
TIA FO 2.2.1 is introducing an improved launch method and RML fiber
bandwidth to achieve MM fiber of 380 MHz-km. Those techniques will not
increase the transceiver cost (or negligible increase). At the 8B/10B
data
rate of 3.125 Gbps, the required fiber minimum bandwidth is about 2.5 GHz.
It is 80% of the bit rate, if all other bandwidth is scaled from the GbE
specification."
A 4 WDM solution for 10 gig throughput as it has been proposed so far
requires ribbon cable. If you are proposing 4 separate 2.5 gig links with
4x the physical density you can use installed MMF. However, for a 10 gig
link, you need ribbon cable, and this is not installed MMF. Sure, it may
be possible to hand tune equal lengths of equal bandwidth cable, but noone
would propose that. There are sketchy skew details in the 4-WDM proposal
so far, but the solutions imply some sort of skew requirement, meaning MMF
fiber has the same delay within the tolerance.
Kevin Demsky
Mixed Signal and VLSI Development
IBM Corp.
3605 Hwy 52 N
Dept. QXS Bldg. 050-2
Rochester, MN 55901
Internal E-mail: kdemsky@ibmusm07
External E-mail: kdemsky@xxxxxxxxxx
>----- Original Message -----
>>
>> Hello Edward,
>>
>> First, thank you for participating in the PAM-5
>> discussion.
>>
>> And now to some of your points:
>>
>> 1) Achievable BERs using PAM-5
>>
>> I would not link multilevel coding with high BERs. The
>> multilevel coding in Fast Ethernet and 1000BASE-T
>> (PAM-3 and PAM-5, respectively) reached
>> only 10^(-10) because the medium, unshielded
>> twisted pair Copper wire, had very strong ISI at
>> the distances they wanted to reach, 100 meters.
>> At 100 meters of cat-5 Copper wire the eye at the
>> input of the receiver is completely closed due
>> to strong ISI.
>>
>> In the present case, we use fiber instead of
>> Copper wire and the target distance of my proposal
>> (4-WDM at 1.25 Gbaud using 850 nm lasers on
>> multimode fiber) is 160 meters using the 160 MHz*km
>> type MMF.
>>
>> At 160 meters using this type of MMF the ISI
>> is small and the optical eye is completely open. Hence,
>> we are back to achievable 10^(-12) system BERs.
>>
>> ---> Personally, I believe that a wide open
>> optical eye at the targetted link length
>> is necessary to reach a BER of 10^(-12).
>> This is another reason why I look with
>> suspicion towards proposals that claim
>> to achieve large link lengths using DFEs
>> (equalizers) to open the eye.
>>
>> 2) Optical links using multilevel coding
>>
>> If there is a need to transport many Gbps
>> on a single limited-bandwidth fiber, and if
>> it is technical feasible and makes economical
>> sense (cheaper) - it will happen. There are
>> many advantages of running at 1.25 Gbaud instead
>> of 3.125 Gbaud.
>>
>> 3) The key issue
>>
>> The big "if", in my opinion, is in the laser
>> linearity. When I was a kid I was puzzled by
>> the kinks in the light-vs-current characteristics
>> of the lasers (I even published two papers in
>> the IEEE J. of QE back in ... xxxx). The problem
>> then was that the lasing filament in the early
>> lasers did not like to stay in the same place
>> because it was not well confined in the
>> lateral dimension (parallel to the layers).
>> The sudden jumps of the lasing filament in the
>> lateral direction as we increased the input
>> current gave rise to the coarse non-linearities
>> in the L-vs-I characteristics (kinks).
>>
>> The solution ? Use lasers with good confinement
>> in the lateral dimension so the lasing filament
>> can not move and then you will achieve good
>> linearity characteristics. I think that we will
>> hear a presentation on laser linearity in the March
>> meeting.
>>
>> 4) technical specifications
>>
>> I agree with you that there is a large amount
>> of work to be done before we can fill-in the
>> Tables with actual numbers in a PAM-5 proposal.
>> In my opinion, this will happen only if we
>> eliminate the present paralyzing multiplicity
>> of PAM-5 proposals and we come out from the
>> March meeting with one and only one clearly
>> supported proposal for PAM-5.
>>
>> Jaime
>>
>> Jaime E. Kardontchik
>> Micro Linear
>> San Jose, CA 95131
>> email: kardontchik.jaime@xxxxxxxxxxx
>>
>>
>> Edward Chang wrote:
>>
>> > Jaime:
>> >
>> > The analyses and data you (and some of Oscar's) presented are very
>> > entertaining, and are very solid engineering work. I really enjoyed
> going through all of them. I believe it is a viable technology.
>> >
>> > However, to implement the PAM-5 coding in an optical link is a new
>> > application, which I am not quite accustomed to it. The
straightforward
>> > digital coding is much more familiar to me. As a result, there are
> several questions in my mind, which I like to address. Furthermore, I
do not
> see many discussions on reflector concerning PAM-5 over an optical
link, and
> I like to initiate one to further enhence its viability.
>> >
>> > In the past, the multiple voltage-level coding was adopted by two LAN
>> > standard, ATM and Ethernet. Both of them were twisted-pair
> applications, and the BER were 10^-10. I proposed BER of 10^-12 in
both working
> groups to be consistent with LAN optical links' BER; however, for some
reason,
> they remain 10^-10, officially in both standards.
>> >
>> > The BER of 10^-12 by a multi-level coding technique has not been proved
> in a real application, or adopted by any standard yet. We need a
vigorous
>> > effort to prove it.
>> > Worst, when the data rate increases from 1.0 Gbps to 10 Gbps, the BER
> should improve to 10^-13 to maintain the 10x advantage in throughput.
>> >
>> > The BER analysis in your presentation is based on the S/N ratio at the
> TIAinput. The system BER will include all other components in a link.
If
> the system BER is 10^-12, the component BER, at least, for the
development
>> > analysis, should be better, or 10^-13 and less.
>> >
>> > It seems, the advantage of the multilevel coding is that the baud rate
> is one half of the digital coding, while maintain the same data rate
for
> both. However, there is an extra price to pay. Not only the "timing
jitter",
> but also the "amplitude jitter" will affect the link reliability, or
BER.
>> >
>> > For the part of the amplitude jitter, there is no specification
existing
> to guide component vendors and system designers to develop a reliable
> products to meet the specified BER.
>> >
The overshoot and ringing of a laser are not characterized, nor
> specified in a specification. They are wide variety of amplitudes and
waveforms. The
>> > maximum allowed external noises (power noise, cross-talk, etc) are not
>> > specified. The tolerance of the signal amplitude is not specified.
In
> a straightforward digital coding, only timing information is important,
> and it does not require amplitude information. The minimum amplitudes
are for
>> > component functionality only, but not data bit information.
>> >
>> > A commonly used equalizer is used to minimize the timing errors caused
> by insufficient bandwidth, which is predictable. For a random, non-
>> > characterized amplitude distortion, it will be hard to compensate.
>> >
>> > The laser output power tolerance is around 8 dB in GbE specification
for
> a low cost source. To tighten it output tolerance will increase cost.
>> > Another uncomfortable reality is that the extinction ratio of the
> narrowest pulse is quite larger (more dc component) than the widest
pulse (less dc
>> > component) - near 1 dB in many cases, which is pattern dependent. This
> will be a very expensive problem to resolve for a level detecting
coding.
>> >
>> > For the timing part, the PAM-5 coding has maximum of three transitions
> in period defined by its baud rate. For example, at the 1.25 G Baud
data
> rate 800 ps period), within an 800 period, there are maximum
transitions of
>> > every 266 ps, which is equivalent of 3.7 G Baud of a NRZ coding. These
>> > pulses are so crowed together to be near sine waves. The advantage is
> that the rise time is slowed down to a sine waveform, which requires a
lower
>> > bandwidth than a pulse shape. However, it added additional disturbing
>> > factor "Crowding Effect". When two pulses of opposite polarity with
>> > different amplitude, or width crowded together (super impose each
> other), the timing (the location of the peak) and amplitude will be
altered.
> The larger one will dominate the smaller one. This is neither the ISI
> caused by bandwidth deficiency, nor the amplitude jitter discussed in
the above
>> > paragraph. It is a third disturbing factor can be minimized by
>> > pre-compensation requiring advanced knowledge of the data pattern.
>> >
>> > I believe, the channel-to-channel skew will be generated not only from
> fiber due to the wavelength difference, but mostly from chips, and PC
board.
>> > Therefore, we will need de-skewing any way, and the fiber length will
be more likely
constrained by the fiber bandwidth.
>> >
>> > One last comment: I believe the bandwidth of an optical link is
> determined by the system rise time equation, 0.8 T = (tT^2 + tF^2 +
tR^2)^0.5 to
> assure a reliable performance. Although the Nyquist theorem defines a
filter
>> > bandwidth related to the bit rate of a NRZ data, for the PAM-5 is not
> NRZ data, which has a maximum of three transitions per a period, the
Nyquist
>> > bandwidth can be adjusted to 3/2T.
>> >
>> > There is another way to achieve an 850 nm, 150 meter installed MM fiber
>> > solution without those issues mentioned above by using 8B/10B-4WDM
> method. TIA FO 2.2.1 is introducing an improved launch method and RML
fiber
>> > bandwidth to achieve MM fiber of 380 MHz-km. Those techniques will not
>> > increase the transceiver cost (or negligible increase). At the 8B/10B
> data rate of 3.125 Gbps, the required fiber minimum bandwidth is about
2.5
> GHz. It is 80% of the bit rate, if all other bandwidth is scaled from
the GbE
>> > specification.
>> >
>> > Therefore, the fiber length L is:
>> >
>> > L = (380x1000)/2500 = 152 meter.
>> >
>> > The 2.5 Gbps VCSEL transceivers are already available from Fibre
Channel
>> > products, and 4 .0 Gbps VCSEL transceivers are coming some time.
>> >
>> > The 8B/10B coding technique is a field proven optical link coding
> technique with multiple vendors supplying all parts to drive the cost
down and to
>> > maintain availability high. We can have a cost-effective 8B/10B-4WDM
> link today, if market is ready.
>> >
>> > Regards,
>> >
>> > Edward S. Chang
>> > NetWorth Technologies, Inc.
>> > EChang@xxxxxxxxxxxxxxxx
>> > Tel: (610)292-2870
>> > Fax: (610)292-2872
>> >
>> >