LVDS problem - Black magic anyone?

I have an LVDS related issue that drives me crazy:

There are two boards with a FPGA that are connected by a ca. 30cm cable. Only 6 wirea are used:
GND + Power
LVDS (with embedded clock), 720Mbps
UART (Rxd + Txd)
(The cable is unshielded for flexibility reasons)

The cable is a "flat Ethernet cable" with 4 twisted pairs, one pair is unused, one pair is LVDS, one pair is GND + Rxd and the last is Power + TxD.

Most cables work fine with a low error rate. However, some cables of the same batch have an excessive error rate. These cables have no visible difference to the good cables.

The end of the cables have the outer isolation removed for about 5cm, the LVDS pair is twisted also for this 5cm, the other wires are straight crimped to the connector.

Now the strange behavior: When touching the cable at some positions by hand, the bad cables are suddenly perfect or at least much better. And the even stranger thing: When touching at the last 5cm, it is NOT the touching of the LVDS pair that makes the difference, but the touching the GND/Power/Uart wires! This can be observed on both sides of the cable.

It is really only "the hand" that makes the difference, strapping together the wires does not help. It does also not look like a mechanical problem of the cable.

I tried changing and even removing the termination resistors, this did not change the behavior at all. I have changed both boards, it is really the cable that makes the difference.

Has anybody an idea what the reason of this behavior could be?

Regards,

Thomas

 

[Theo Markettos]

thomas.entner99@gmail.com wrote:

I have an LVDS related issue that drives me crazy:

There are two boards with a FPGA that are connected by a ca. 30cm cable.
Only 6 wirea are used:
GND + Power
LVDS (with embedded clock), 720Mbps
UART (Rxd + Txd)
(The cable is unshielded for flexibility reasons)

The cable is a "flat Ethernet cable" with 4 twisted pairs, one pair is
unused, one pair is LVDS, one pair is GND + Rxd and the last is Power +
TxD.

Most cables work fine with a low error rate. However, some cables of the
same batch have an excessive error rate. These cables have no visible
difference to the good cables.

What category cable is it? Even cat6 or cat6a aren't intended for use
beyond 500MHz. While there's less loss because the cable is shorter, it's
worth looking up the loss curves and doing the calculation. You have
checked that the 'flat' cable does actually contain twisted pairs?

What's the position of the ground with respect to the LVDS pair? Is it
adjacent, or further away? Are these direct electrical connections, or is
there any signal isolation? Is the spare pair floating?

What's the grounding arrangements of both boards? Is this the only ground
connection between them? How are they powered? Do they come from the same
AC source, or different sources? What's the arrangement of earths, and any
possibility they could be on different AC phases?

What are you driving the RXD/TXD signals with? Do you know their rise time?

For that matter, do you know the rise time of the LVDS signals?
What are you using for an LVDS receiver? (Any equalisation?)
What's the encoding on the wire (8b/10b, Manchester, etc)?

I tried changing and even removing the termination resistors, this did not
change the behavior at all. I have changed both boards, it is really the
cable that makes the difference.

Have you looked at the signal on a scope?

Where did the 'termination resistors' come from? How did you decide what
topology and what values to use?

> Has anybody an idea what the reason of this behavior could be?

Some ideas:
1. You're changing the capacitance of the cable by touching it
2. You're creating a capacitive path to ground through your body
3. You're changing the topology of the inductive loop between an LVDS line
and ground
4. You're coupling other signals into the LVDS pair (eg RXD), especially if
the twists have ended up aligned
5. The signal has higher frequency components than you think and these are
messing up the receiver. Extra capacitance by touching increases their
loss, enough for the receiver to cope
6. The cable is simply too lossy at the frequency you're using
7. You have a grounding issue with the setup
8. The line isn't terminated correctly so you're getting reflections which
interfere with the signal

Do you have a network analyser available? Looking at the insertion loss of
the cable over a range of frequencies may tell you something about what's
going on.


I've done a lot of work with SATA cables. Bottom line is, some are just
cheap and simply don't work. SATA error recovery is enough to mask many of
the problems when used in a PC, so nobody notices.

Theo

 

I have an LVDS related issue that drives me crazy:

There are two boards with a FPGA that are connected by a ca. 30cm cable..
Only 6 wirea are used:
GND + Power
LVDS (with embedded clock), 720Mbps
UART (Rxd + Txd)
(The cable is unshielded for flexibility reasons)

The cable is a "flat Ethernet cable" with 4 twisted pairs, one pair is
unused, one pair is LVDS, one pair is GND + Rxd and the last is Power +
TxD.

Most cables work fine with a low error rate. However, some cables of the
same batch have an excessive error rate. These cables have no visible
difference to the good cables.


What category cable is it?

CAT5e (UTP)

Even cat6 or cat6a aren't intended for use
beyond 500MHz. While there's less loss because the cable is shorter, it's
worth looking up the loss curves and doing the calculation. You have
checked that the 'flat' cable does actually contain twisted pairs?

I have opened up a cable now - the pairs are twisted but the turn rate is much lower than what I would have expected - it is about one full turn every 20cm or so.

> What's the position of the ground with respect to the LVDS pair?

Ground is (only) in this cable, it is paired with RxD:
1st pair: GND + RxD
2nd pair: LVDS
3rd pair: Power + TxD
4th pair: unused

Is it
adjacent, or further away? Are these direct electrical connections, or is
there any signal isolation?

Ground is only connected through the cable between the board. The transmitter side (a small sensor board) has no other connections to it at all. The receiver side has other connections like HDMI, etc.

In fact, there is a ferrite bead on the sensor PCB in the ground and in the power path (directly at the connector). But I also have already bridged this beads with zero ohm -> no change.

> Is the spare pair floating?

Yes. I have removed this wires from one bad cable now -> no change.

What's the grounding arrangements of both boards? Is this the only ground
connection between them?
See above

How are they powered? Do they come from the same
AC source, or different sources? What's the arrangement of earths, and any
possibility they could be on different AC phases?
The receiving board is powered via an external +12V power supply, the sensor board is only powered via this cable.

What are you driving the RXD/TXD signals with? Do you know their rise time?
With 3.3V FPGA I/O pins with series resistors and even some small cap at the connector. The rise-time is very slow. I have also stopped the UART communciaton for testing -> no change.

For that matter, do you know the rise time of the LVDS signals?
Unfortunately: No (I have only a 500 MHz scope)

What are you using for an LVDS receiver? (Any equalisation?)
It is a Cyclone IVGX SerDes. Yes, equalization is active (at a medium setting). I will play around a little bit with this settings (unfortunately the compile-time is quite long).

What's the encoding on the wire (8b/10b, Manchester, etc)?
It is quite a simple protocol with start- and stop-bit (10b data-word) as used by DS92LV1023 (no DC balancing)

I tried changing and even removing the termination resistors, this did not
change the behavior at all. I have changed both boards, it is really the
cable that makes the difference.
Have you looked at the signal on a scope?
Yes (as good as possible with a 500MHz scope...): I see no visible change in all the signals when touching the cable. Only the dramatic effect in the error rate. What surprises me is that the scope probe has almost no influence on the error rate while the touching by hand does...

Where did the 'termination resistors' come from? How did you decide what
topology and what values to use?
They are 100R on both sides (as this is what I expect as impedance of a CAT5e cable). I put a 100R also on the transmitter side as this improves the robustness in my experience, esp. with longer cables. I also removed the 100R on the transmitter side and it did not help.

Has anybody an idea what the reason of this behavior could be?
Some ideas:
1. You're changing the capacitance of the cable by touching it
2. You're creating a capacitive path to ground through your body
3. You're changing the topology of the inductive loop between an LVDS line
and ground
4. You're coupling other signals into the LVDS pair (eg RXD), especially if
the twists have ended up aligned
5. The signal has higher frequency components than you think and these are
messing up the receiver. Extra capacitance by touching increases their
loss, enough for the receiver to cope
6. The cable is simply too lossy at the frequency you're using
7. You have a grounding issue with the setup
8. The line isn't terminated correctly so you're getting reflections which
interfere with the signal
I will have to think further about these...

Do you have a network analyser available? Looking at the insertion loss of
the cable over a range of frequencies may tell you something about what's
going on.
Unfortunately: No.

I've done a lot of work with SATA cables. Bottom line is, some are just
cheap and simply don't work. SATA error recovery is enough to mask many of
the problems when used in a PC, so nobody notices.

Theo

Thanks a lot. This discussing this issue with someone really helps. I will post new results here when I have any.

Regards,

Thomas

 

What are you using for an LVDS receiver? (Any equalisation?)
It is a Cyclone IVGX SerDes. Yes, equalization is active (at a medium
setting).
I will play around a little bit with this settings (unfortunately
the compile-time is quite long).

It turned out that the equalization was too high. When increasing the setting, things got worse. When setting it to "low", things work perfect without any errors...

I think the great effects of touching the cable can be explained with the weak twisting of the pairs.

Regards,

Thomas

 

[Theo Markettos]

thomas.entner99@gmail.com wrote:

It turned out that the equalization was too high. When increasing the
setting, things got worse. When setting it to "low", things work perfect
without any errors...

I think the great effects of touching the cable can be explained with the
weak twisting of the pairs.

I suspect the cable looks more like a parallel ribbon cable than twisted
pair. That means there could be other signals between the twisted pair and
ground - eg if it turns out it goes:
GND RXD D+ D- TXD VCC

Any common mode currents, which there will be because your coding scheme is
not DC balanced, will flow around a loop with RXD or TXD in the middle, so
are likely to be inductively coupled with these wires (ie cause crosstalk).
It's difficult to say what effect the inductance of this loop will have on
the LVDS signal (I suspect not a huge amount, but have no numbers: it will
depend on how quick the rise times are), but I wouldn't be surprised if you
saw coupling into the RXD and TXD wires.

Likewise, we'd have to know a bit more about the channel characteristics to
predict what the likely frequency response of the channel might be (you'd
model each part's LCR properties separately: hard numbers needed).

I'd guess that the touching effect increases both the capacitance to ground
and the inter-pair capacitance. It moves the frequency nulls around - I'm
not familiar with the equalisation settings on that transceiver to know the
effect of the different modes, but it's not implausible that equalisation
could cope with one but not the other. In theory your transmission line
should be correctly terminated so you don't need equalisation, but it very
obviously isn't.

Theo

 

[glen herrmannsfeldt]

thomas.entner99@gmail.com wrote:

I have an LVDS related issue that drives me crazy:

There are two boards with a FPGA that are connected by a ca.
30cm cable. Only 6 wirea are used:
GND + Power
LVDS (with embedded clock), 720Mbps
UART (Rxd + Txd)
(The cable is unshielded for flexibility reasons)

The cable is a "flat Ethernet cable" with 4 twisted pairs,
one pair is unused, one pair is LVDS, one pair is GND + Rxd
and the last is Power + TxD.

Ethernet cables use pin pairs (1,2) (3,6) (4,5) and (7,8).

With the usual cables, you can see the pairs, including the
colors, in the connector.

I have one, though, which I would call flat, and which the pairs
are not visible. It is very thin and flexible, and seems to work
fine for 100baseTX. I suspect the wire is much smaller than the
normal Cat 5 wire.

But otherwise, ethernet cables are normally rated to 100MHz.
1000baseT uses a complicated signal system on all four pairs
(in both directions at the same time) to get the bits through.

I wouldn't be surprised if you could get 720MHz through Cat 5
cable for 30cm, but your cable might not be quite that good.

-- glen

 

Den onsdag den 13. august 2014 23.02.24 UTC+2 skrev Jon Elson:

thomas.entner99@gmail.com wrote:





The cable is a "flat Ethernet cable" with 4 twisted pairs, one pair is

unused, one pair is LVDS, one pair is GND + Rxd and the last is Power +

TxD.

This is NOT LVDS! The D stands for Differential! You use one wire of the

pair for the true signal, one wire for the complement signal. I don't even

know how you do single-ended with LVDS receivers, but it won't work well,

as you clearly have discovered.



You can still use LVDS the right way with 3 pairs. One pair is power &

ground, one is the Rxd true/compl pair, the other is the Txd true/compl

pair.

afaiu, he is using one pair for LDVS, the other rx/tx is single ended uart
paired with gnd and power. So the wiring is ok

-Lasse

 

[Theo Markettos]

Jon Elson <jmelson@wustl.edu> wrote:

This is NOT LVDS! The D stands for Differential! You use one wire of the
pair for the true signal, one wire for the complement signal. I don't even
know how you do single-ended with LVDS receivers, but it won't work well,
as you clearly have discovered.

You can still use LVDS the right way with 3 pairs. One pair is power &
ground, one is the Rxd true/compl pair, the other is the Txd true/compl
pair.

The OP stated he has one LVDS pair (D+ and D- twisted), and two UART signals
(RXD and TXD). The UART signals aren't sent by LVDS, they're just other
things down the same cable. I assume the OP's LVDS is unidirectional (or
else there would be other issues). The OP's situation of having RXD+GND and
TXD+VCC pairs, rather than RXD+TXD and GND+VCC, is probably slightly better
but it's marginal at typical UART speeds (and as always depends on the setup
- as we've discovered the 'twisted pair' isn't).

You can connect single-ended into an LVDS receiver by grounding the negative
input (if you have positive and negative voltage rails) or putting it
mid-range via a resistive divider (on a single supply arrangement) but
you've just halved your signal amplitude (and thus differential SNR) - and
of course you have no protection from common-mode noise.

Theo

 

[Jon Elson]

thomas.entner99@gmail.com wrote:

The cable is a "flat Ethernet cable" with 4 twisted pairs, one pair is
unused, one pair is LVDS, one pair is GND + Rxd and the last is Power +
TxD.

This is NOT LVDS! The D stands for Differential! You use one wire of the
pair for the true signal, one wire for the complement signal. I don't even
know how you do single-ended with LVDS receivers, but it won't work well,
as you clearly have discovered.

You can still use LVDS the right way with 3 pairs. One pair is power &
ground, one is the Rxd true/compl pair, the other is the Txd true/compl
pair.

Jon

 

[alb]

hi Jon,

Jon Elson <jmelson@wustl.edu> Wrote in message:

Well, one other thing, is he terminating the LVDS with a resistor matching
the characteristic impedance of the cable? I'll bet he isn't, his results
would VERY likely match what he describes. We use 100 Ohm resistors, they
are close enough. For best results, tune between 110 and 120 Ohms for
minimum reflection.

from the OP:
I tried changing and even removing the termination resistors,
this did not change the behavior at all. I have changed both
boards, it is really the cable that makes the difference.

BTW, removing the terminators on an LVDS - single drop - line is
*never* a good idea. LVDS drivers are current sources and the
receiver may simply not tolerate the differential swing on it's
big input impedance.

I did not understand how the op solved his problem changing
receiver equalization, he clearly mentioned some cables did work
ok without the need to change equalization...

Al

 

Wow, quite some discussion going on here... (I have considered this myself "solved" and working on other stuff again.)

To give some comments:
Yes, we use 100 ohm termination, in fact on both sides of the cable. I tried to remove the termination on both sides (but I don't think I have removed both sides at the same time - would not really make sense...). I have also checked with a multimeter and got the correct values (e.g. 50 ohm, when both are mounted and cable connected.)

My conclusion was, that with the original (quite strong) equalizer setting things were barely working (even with "good" cables I got an transmission error here and then while "bad" cables resulted in errors all over). I think it was simply to strong for the short cable distance. With the lowest possible equalizer setting we have not seen any transmission errors anymore with any combination of boards and cables.

I had dismantled one cable and found that the (while they were really twisted) the turn rate was very low, about 1 to 2 full turns about the complete length. This might be good enough for long ethernet cables, but I can imagine that this short length make some of them almost behave like untwisted cables with quite some cross talk. This is my explanation why touching the untwisted wires had some influence although I am still surprised that the effect was such a drastic improvement.

What I still cannot understand is why the (in my opinion) much more dramatic change of removing the termination resistors had no influence (not good nor bad) on this obviously barely working system while the touching had.

Regards,

Thomas

 

[Jon Elson]

Theo Markettos wrote:

The OP stated he has one LVDS pair (D+ and D- twisted), and two UART
signals
(RXD and TXD). The UART signals aren't sent by LVDS, they're just other
things down the same cable.
OK, he confused me with the UART data signals.

Well, one other thing, is he terminating the LVDS with a resistor matching
the characteristic impedance of the cable? I'll bet he isn't, his results
would VERY likely match what he describes. We use 100 Ohm resistors, they
are close enough. For best results, tune between 110 and 120 Ohms for
minimum reflection.

Jon

 

[alb]

Hi Thomas,

thomas.entner99@gmail.com wrote:
[]

To give some comments: Yes, we use 100 ohm termination, in fact on both
sides of the cable. I tried to remove the termination on both sides (but I
don't think I have removed both sides at the same time - would not really
make sense...). I have also checked with a multimeter and got the correct
values (e.g. 50 ohm, when both are mounted and cable connected.)

I do not understand why you need termination at both ends. There are two
aspects when terminating an LVDS line:

1. line impedence matching.
2. provide a load for a current source.

In both cases it does not make too much sense to have a termination at the
source.

My conclusion was, that with the original (quite strong) equalizer setting
things were barely working (even with "good" cables I got an transmission
error here and then while "bad" cables resulted in errors all over). I think
it was simply to strong for the short cable distance. With the lowest
possible equalizer setting we have not seen any transmission errors anymore
with any combination of boards and cables.

Did you see your signal on the scope *after* the receiver for both a /good/
and /bad/ cable? Do you have an eye diagram?

I had dismantled one cable and found that the (while they were really
twisted) the turn rate was very low, about 1 to 2 full turns about the
complete length. This might be good enough for long ethernet cables, but I
can imagine that this short length make some of them almost behave like
untwisted cables with quite some cross talk.

twist per meter is usually specified in the cable datasheet; considering the
rate you're working at I'd seriously consider the cable as a fundamental part
to be analyzed.

This is my explanation why
touching the untwisted wires had some influence although I am still
surprised that the effect was such a drastic improvement.

I would not call 'improvement' something that you cannot justify or explain.
Since you do not know what is going on, it is *very* possible the problem is
simply hidden for a various combination of causes and will likely reappear
later on.

What I still cannot understand is why the (in my opinion) much more dramatic
change of removing the termination resistors had no influence (not good nor
bad) on this obviously barely working system while the touching had.

IMO your connection is far from being a proper one, with serious matching
issues and a very scattered amount of information to analyze. With these basis
any conclusion on 'working' or 'not working' are pointless at best.

If you don't have a model to explain why it works and why it doesn't, your
likely exposed to see the issue again later and thanks to Murphy's law it
would be right two days before shipping!

If you want to seriously remove your issues you need to be able to explain
every detail, spot the cause and remove it. Otherwise it would just be another
piece of guesswork.

Al

 

[Tom Gardner]

On 17/08/14 17:57, alb wrote:

Hi Thomas,

thomas.entner99@gmail.com wrote:
[]
To give some comments: Yes, we use 100 ohm termination, in fact on both
sides of the cable. I tried to remove the termination on both sides (but I
don't think I have removed both sides at the same time - would not really
make sense...). I have also checked with a multimeter and got the correct
values (e.g. 50 ohm, when both are mounted and cable connected.)

I do not understand why you need termination at both ends. There are two
aspects when terminating an LVDS line:

1. line impedence matching.
2. provide a load for a current source.

In both cases it does not make too much sense to have a termination at the
source.

See "LVDS Owner's Manual Design Guide , 4th Edition"
http://www.ti.com/lit/ml/snla187/snla187.pdf

Did you see your signal on the scope *after* the receiver for both a /good/
and /bad/ cable? Do you have an eye diagram?

Ensure the probes don't disturb the signal - which is easier said
than done.