>
> I see in XAPP475 they say that "Unfortunately, IBIS 3.2
> still is not widely supported by simulators.".
>
Oops, I missed that explanation when skimming XAPP475
yesterday to find the TLine parameters.
I was actually thinking of this sentence from Answer Record
#19320 when I griped about Xilinx still using IBIS 2.1:
"We do not have an IBIS model for LVDS_25_DT, as
the IBIS specification does not provide a mechanism
for representing the true differential termination"
It's been I while since I read the IBIS specs, but I believe
it was IBIS 3.x which added support for modeling a differential
terminator, instead of having to bury single-ended terminator
currents in the GND/VCC clamp table.
> It wouldn't hurt to publish new IBIS3.2 or even 4.0 files
> alongside the old ones though!
My interest in having data available in one of the newer
IBIS versions is not to actually use them in an IBIS simulator,
but to have {almost} human-readable documentation of the
differential package and I/O parasitics, which Xilinx doesn't
currently publish in any other form.
I suspect if you use the differential I/O standards, and then
tie adjacent I/O pins as strong drivers to GND(VCCO), you can
establish a nice GND S+ S- GND(VCCO) pinout in a leaded package
and have a good shot at doing extremely fast I/O in an el-cheapo
package.
And I'd love to see a small S3 sold in one of the enhanced
VQFP ground-paddle packages, for both thermal and electrical
reasons, especially if done with a G S+ S- G pinout.
>
>Thanks for a very informative post, much appreciated
>
As were yours, thanks - I've been building 4-6 layer boards
with FPGA's powered from localized plane fills, as you describe,
for many years and generations of FPGA. Keeping the 'dirty' FPGA
power plane localized in such a fashion also helps in RF/mixed
signal board layouts.
I took a look at building a simple first or second order SPICE
package model of the leaded S3 parts a few months ago, but there
wasn't any package data available in the IBIS files; if I get a
chance to take another crack at it, I'll post some LTspice files.
And if anyone out there has built a test fixture and made
either differential TDR or VNA measurements on some of the S3
leaded packages, I'd love to see some real world data to help
model the pin-pin coupling for a G S+ S- G pinout :)
Brian
Reply by Brian Drummond●October 21, 20042004-10-21
On Tue, 19 Oct 2004 13:51:49 -0700, "Symon" <symon_brewer@hotmail.com>
wrote:
>"glen herrmannsfeldt" <gah@ugcs.caltech.edu> wrote in message
>news:cl3t0g$qob$1@gnus01.u.washington.edu...
>>
>> The suggestion was to improve the ground plane by making it
>> thicker. To reduce via inductance you need as many of them
>> as you can get. (Inductors in parallel.) The OP wanted a four
>> layer board instead of following the suggested two ground and
>> two power planes.
>>
>I just realised, making it thicker will only reduce its resistance. Its
>inductance won't change. The inductance is determined by the loop area of
>the current path. Think about it, when you calculate the inductance of a
>coil, you don't need to know the wire diameter, just the diameter and number
>of turns.
What you can do, though, is play with the thickness of the prepreg. Make
it thin enough and you increase the capacitance between ground and power
planes. Not to a very high value, granted, but connected to the planes
by a _very_ low inductance!
- Brian
Reply by Austin Lesea●October 20, 20042004-10-20
All,
IBIS 4.0 is now out for bid for the "golden parser."
Without a "golden parser", no one can say they adhere to the standard.
The last golden parser was back in the IBIS 1.2 days, so that is all
anyone can lay claim to.
If you claim something more than that, you are not being entirely
honest. Small errors in the IBIS file get rejected by various tools,
and support mushrooms.
As well, not everyone implemented the changes in IBIS 2, 3.... the same
way, or the simulators implementations were also not identical....
We have 200,000+ seats of software out there, and trying to blaze a new
trail for IBIS is like pushing tons of wet spaghetti (hard work, and not
very satisfying).
We are actively looking at when to fold in the new IBIS, but only after
it is supported by the tools that allow us to succeed. Until then, we
use what we have got, which includes the encrypted hspice versions for
folks that have to have the "answer."
For the MGTs, no one can say if IBIS is adequate or not for differential
signals at 10 Gbs, so we are sticking to methods that we know work. We
do know that behavioral models are really fast, but also really
inaccurate. If someone has a complete backplane simulation with
extracted pcb parasitics (s parameters, complex lossy t-lines,
connectors, pre-emphasis transmitter and adaptive receiver, etc.) that
is within 5% of the behavioral model, I'd like to know how the miracle
occurred.
Austin
Reply by Symon●October 20, 20042004-10-20
"Brian Davis" <brimdavis@aol.com> wrote in message
news:a528ffe0.0410200342.638dedb0@posting.google.com...
<snipped very informative stuff>
>
> The other concern I have with the Xilinx IBIS models is
> that they're still using an ancient version of IBIS (2.1),
> which doesn't support some of the newer IBIS features such
> as differential input parasitics. ( Which explains the lack
> of IBIS models for the _DT terminators in the V2Pro )
I see in XAPP475 they say that "Unfortunately, IBIS 3.2 still is not widely
supported by simulators.". It wouldn't hurt to publish new IBIS3.2 or even
4.0 files alongside the old ones though!
Thanks for a very informative post, much appreciated.
Best, Syms.
Reply by Brian Davis●October 20, 20042004-10-20
Symon wrote:
>
>I wonder, do the IBIS models include the lead frame?
>
AFAIK, IBIS modeling ignores ground bounce & such.
Earlier versions of the S3 IBIS files used the SAME package
model for ALL packages, from BGA through PQFP; the IBIS file
had tiny entries for package parasitics, and Xilinx suggested
a user-added Tline model to model the pin:
uncoupled 65 ohm transmission line, 25-100 ps delay
from :
http://direct.xilinx.com/bvdocs/appnotes/xapp475.pdf
Now, if what you're modeling is a transmission line, or
behaves like one at the edge rates of interest, a Tline
model is more appropriate than a single element lumped
approximation; but to model a conventional leadframe package
with an UNCOUPLED** Tline model seems rather optimistic.
Looking again today, the latest version (2.6) of the S3 IBIS
files now has lumped parasitics for some of the leaded packages
( PQ208 and TQFP144, but not the VQ100 ). Note: XAPP475 has not
yet been updated to address this modeling change.
The other concern I have with the Xilinx IBIS models is
that they're still using an ancient version of IBIS (2.1),
which doesn't support some of the newer IBIS features such
as differential input parasitics. ( Which explains the lack
of IBIS models for the _DT terminators in the V2Pro )
** I don't believe IBIS 2.1 supports modeling of pin-pin coupling
Brian
Reply by Hal Murray●October 20, 20042004-10-20
>That's because the lead frame has buggered your SI before you've even
>started your PCB.
Then why does Xilinx sell the chips? They must be good for something.
Maybe there is a comment in some ap-note that says not to bother
trying to implement PCI in a PQ-208 type package. (and explaining
why) Or maybe it takes 8 layers, or ...
I haven't seen anything like that, but I haven't looked carefully
and nobody has mentioned anything like that yet in this discussion.
"Don't do that" might be the right answer.
--
The suespammers.org mail server is located in California. So are all my
other mailboxes. Please do not send unsolicited bulk e-mail or unsolicited
commercial e-mail to my suespammers.org address or any of my other addresses.
These are my opinions, not necessarily my employer's. I hate spam.
Reply by Symon●October 20, 20042004-10-20
"Brian Davis" <brimdavis@aol.com> wrote in message
news:a528ffe0.0410191907.7dea9fd4@posting.google.com...
> Symon wrote:
>
> < re. the missing leaded package SSO data >
>>
>>That's because the lead frame has buggered your SI before
>>you've even started your PCB.
>>
>
> Yes, and the poorer on-die power distribution of those
> same leadframes makes adhering to those (missing) SSO
> guidelines even more important than in the BGA parts.
>
V. good point. I wonder, do the IBIS models include the lead frame?
>>
>>The idea is that if the voltage mode drivers switch
>>simultaneously in opposite directions, the current through
>>the Vcco pins stays constant, so the lead/trace inductance
>>doesn't screw things up.
>>
>
> Xilinx has two flavors of differential driver:
>
> 1) the older style, voltage-mode, pseudo-differential,
> switch-two-CMOS-outputs-with-an-external-resistive-attenuator
>
> which they say doesn't balance that well :
> http://www.fpga-faq.com/archives/42700.html#42709
>
> 2) the 'real' balanced current-mode drivers of the V2/V2P/S3,
> which the Answer Records say is not a factor in SSO limits
>
> So, I'd expected to see much better SSO numbers for the 'real'
> LVDS drivers than for the older 'pretend' ones, but DS099, table 23,
> says both types have a four-pair SSO limit per VCCO-GND pair.
>
>
> Brian
I got the difference between the two types, but failed to realise that they
didn't balance so well. Thanks for the link you posted to Bob and Austin's
exchange, I'm thinking again.
I guess I'm very cynical, I think the "25 ps to 125 ps" time from the pad to
the pin is a red herring. The synchronicity of the IOBs outputs switching on
the die is what's important.
Maybe the 'gate' or drive capacitance of the output transistors added to the
bounce? Maybe a tiny bit! I guess that leaves the crossover current, but how
long does that last? I thought the output structure stopped that happening.
CMOS, right? And how did they measure it? Hmmm!
Cheers, Syms.
Reply by Brian Davis●October 20, 20042004-10-20
Austin wrote:
>
>LVDS is a low current driver, but the LVPECL is two single
>ended drivers with external resistors, so it has significant current.
>
>Even LVDS has a restriction, but not nearly that of a
>larger driver (more current). Unless I am missing something?
>
I'd also expected to see much better numbers for the
current-mode drivers, but the Spartan3 SSO table that
you referenced says otherwise:
Excerpts from DS099-3, v1.4, Table 23
SSO outputs per VCCO/GND pair
Single ended:
LVCMOS33 Fast 16mA 7 pins
LVCMOS33 Fast 24mA 3 pins
Current-mode:
LVDS_25 4 pairs (8 pins)
RSDS_25 4 pairs
Psuedo-differential:
BLVDS_25 4 pairs
LVPECL_25 4 pairs
Which lists the current mode LVDS drivers as having the
same low SSO pin limit (4 pairs) as the LVPECL drivers, and
pretty much the same pin limit as does LVCMOS33/FAST/16mA.
>
>>
>> 3) Why do the input-only differential parallel DCI standards
>> show up in the SSO table?
>
>Because they draw current for the parallel termiantion, and the
>restriction is for current drain on the Vcco/Gnd pins in the bank.
>
I'd hoped that was the reason; but since those same "4"'s were
next to every type of differential I/O in the table, I thought it
might be a block pasting error when compiling the datasheet.
Following up on this:
In "normal" operation, these on-chip DCI terminator pairs have
already been biased to 1.25 V, and experience a small balanced
input swing (~0.8V), providing some measure of terminator VCCO
current cancellation.
In this "normal" mode of operation, in the BGA packages,
they have a four pair SSO limit, possibly less for leaded parts.
But at post-configuration DCI startup, these terminators all
switch simultaneously, unbalanced, from full stop to 1.25 V
Given that this unbalanced, larger swing should have poorer SSO
limits than the "normal" operation, that's why I was concerned over
on that other thread about using LVDS_25_DCI in the leaded packages.
Brian
Reply by Brian Davis●October 20, 20042004-10-20
Symon wrote:
< re. the missing leaded package SSO data >
>
>That's because the lead frame has buggered your SI before
>you've even started your PCB.
>
Yes, and the poorer on-die power distribution of those
same leadframes makes adhering to those (missing) SSO
guidelines even more important than in the BGA parts.
>
>The idea is that if the voltage mode drivers switch
>simultaneously in opposite directions, the current through
>the Vcco pins stays constant, so the lead/trace inductance
>doesn't screw things up.
>
Xilinx has two flavors of differential driver:
1) the older style, voltage-mode, pseudo-differential,
switch-two-CMOS-outputs-with-an-external-resistive-attenuator
which they say doesn't balance that well :
http://www.fpga-faq.com/archives/42700.html#42709
2) the 'real' balanced current-mode drivers of the V2/V2P/S3,
which the Answer Records say is not a factor in SSO limits
So, I'd expected to see much better SSO numbers for the 'real'
LVDS drivers than for the older 'pretend' ones, but DS099, table 23,
says both types have a four-pair SSO limit per VCCO-GND pair.
Brian
Reply by Symon●October 19, 20042004-10-19
"glen herrmannsfeldt" <gah@ugcs.caltech.edu> wrote in message
news:cl44k8$1qu$1@gnus01.u.washington.edu...
> I think, though, that more vias still decreases the
> inductance until they get so close that there is a
> large field overlap bewteen them.
Agreed, as you said earlier "Inductors in parallel". The current is shared
between the vias. I think you're right about the total inductance increasing
as the two (say) vias get close as well, food for thought. Is that because
of the mutual inductance between the two current loops? Time to crack out a
text book!
An interesting discussion, thanks Glen!
Best, Syms.