Bill,
My comments are in-line. -joel
The channels
submitted in goergen_02_0709
and goergen_02_0904 certainly do have great value and should be used by
the
working group for further work. However, they should not be used to
the
exclusion of other channels that represent complete channels. Both the
signaling work and the setting of the channel limit should be based on
a wide
selection of data and must include data that represents real channels.
I
submit the following reasons to support this conclusion.
I have never advocated using only these channels. They are not
meant to be used as golden. I'm not sure where you or anyone else
would have gotten that as the objectives had been clearly stated in
many phone conference calls. I would recommend taking some time to
review past discussions on this matter.
Your choice of the words 'wide selection' is interesting. What you
imply is that you want the channel limit to include a wide variety of
channel design practices.
1)
A
connector
agnostic test system is a worthy goal. However, this is not achieved
by
omitting the connectors. Additionally, each connector is accompanied
by
two via fields with their associated breakouts. This is not well
represented by a single via transition.
If you actually take the time to employ the research, it is
easily shown that in short segments, the second set of holes you
describe is not seen as a double effect. It's the first set that both
determine and dominate the response, the second being absorbed within
the first. The dominant factor is the first holes, combined with the
short segments. Again, show me what flavors of nets I have not been
able to show in these cards. I have matched everything to data, good
and bad, yours and everyone else's.
2)
A test
system does not necessarily take into account the tradeoffs that
inevitably
occur in the design of a cost-effective, high density system.
What you imply is this test system is too good of a model. It
should be in most respects. It uses over 20 years of research and
design practice in back plane design, combined with proper routing and
geometry shapes. I took the cards further and employed poor design
practice in the aggr signals ... sort of a two for one when including
multi-disturber NEXT/FEXT. It shows the effects of stubs, as well as
the effects of xtalk from passing in and out of poor geometry via
shapes. I clearly identified this in my presentations. It matches
quit well with your data.
A test system DOES take into account trade-offs. Otherwise, what's the
point of creating it. All of the methods employed in these test cards
are used everyday in a 'cost-effective, high density system'. I think
the market speaks well enough that this system implementation is cost
effective. And the data of my work is clearly, in terms of the market,
more dense then ATCA applications.
What you fail to state is that the real issue here is the inclusion of
both gross and severe stub effects within the channel, further
disturbed with one-two combinations of short traces. My data shows
this quit well, as does other public data. Stub effects are very nasty
on the channel. There are cost effective ways to deal with them. You
chose not to deal with them at all in your data. Your solution is to
lower the channel model and let the silicon deal with this problem you
call ' a design trade-off'. I disagree. The channel should not
include severe and gross stub effect. We've gone through this path in
xaui ... one of two major reasons people believe the xaui model is
broken. So instead of attacking my data, why not stick to the issue
... stub effect and short traces.
If you stick to the issue and clearly examine the OIF approach in CEI,
the reason for stateye becomes clear ... you need something very
complex to simulate a system with channels registering effects of gross
design practice termed by you to be trade-offs in a cost effective and
dense design. Stateye is barely used AND understood by those using it
and is clearly unnecessary to simulate just the channel effects (not
discussing the system here with tx and rx as there are benefits) when
good design practice is employed. Stubs are easy to both control and
effectively design around.
Once I understand the protocol for identification, I will show a trace
with stubs, and the same trace with stubs that is influenced by
patented technology.
So, you want stubs, I don't. Stop attacking the validity of my data
and test cases. Start focusing on getting acceptance on the inclusion
of gross and severe stubs into the channel model.
The problem in dealing with both short track effects AND stubs is that
the power and die size are going to be larger then I think we can
include in a telecom design ... our systems are too dense to support
the kind of power required for a 5Ghz drop-off ... I thought it would
be 6Ghz, but when you evaluate the stub and short tracks, the data
shows a falling off the graph before 5Ghz. You can see this on the
public xaui data.
If we are going to include channels such as this, then I would strongly
suggest two channel models and perhaps two sets of phys. It is my
humble opinion that this would happen anyway because the dense systems
employed by the telecom group will not be able to handle the power and
cooling that a fractional dense ATCA system could handle, essentially
making the standard unusable for the telecom market. I'm not even sure
any of the signaling discussed to date can handle that severe channel.
3)
We can
only compare
the public data and there are significant differences. It is difficult
to
predict how these differences will impact the performance for any
particular
line code / equalization as illustrated in sinsky_01_0904.
These test cards show several types of channels, not just one
type. Hence that was one of the reasons for doing the cards. To
purposefully identify several types of conditions in such a manor as to
evaluate without getting lost in the volume of data. Again, I have
never stated that these are replacements for use of anything. They are
channels used as a common point of reference so when two people show
different conclusions of different data, if they both verify similar
results on a set of test cards such as mine, we know the data is
accurate and taken in a similar manor ... calibrated to some extent.
Thus making the conclusions interesting and relevant.
This topic has been with us on several channel discussion in various
bodies since before the days of 802.3z, fiber or copper. How do we
verify everyone is taking data the same way and if they don't, how does
that effect the outcome? We have answered that question after xaui,
CEI, and others, as well as ours. There is a difference, That was one
reason we spent so much time on recommended VNA set-ups.
Again, the test board
data
from goergen_02_0709 and goergen_02_090 should be considered, but must
be put
into perspective. We cannot rely exclusively on incomplete test
systems
or contrived models however valuable they may be.
-Bill Peters
Fellow collegues,
During the back plane meeting this week in Ottawa, it was pointed out on several
occasions by one colleague that the channels submitted for review and
simulation as reported in goergen_02_0709 and goergen_02_0904 are not
real
channels and therefore, should not be used by the working group in any
capacity.
I've studied this concern at length and defend my work in whole as
follows.
1) The intent of a connector-less test card was agreed to in part by
the
members of the channel ad-hoc. It's creation, though not a new concept
at
all, was to prevent a common set of test cards from becoming connector
evaluation boards by the industry. My understanding of the ATCA
platform
is that this MSA has agreed to 'ONE' connector. Therefore, evaluating
comparative back planes in an ATCA environment is easy. It does not
address the diverse nature of the telecom segment, which is the primary
intent
of the connector-less cards.
2) I've examined all the channel data presented to the public, as well
as that
submitted to me in private. All of the public data submitted to IEEE
as
presentations, and all of the private data sent to me directly
correlates to
the data from the connector-less back plane cards. Do the comparison
yourself ... Look at every channel and compare it to all the data
submitted and
you will see the same likeness. The data speaks for itself. All the
effects form various degrees of inferior channel design are included,
certainly
not excluding any of those seen in Peters_01_0904.
Therefore, if my data is not valid to consider, then what data 'IS'
valid for
consideration? It's clearly acceptable to simulate a channel and then
measure the likeness to real before continuing with the remaining
simulations/approximations. These test back planes correlate perfectly
and I stand by the design, implementation, and data collected. If the
majority body feels the data is not valid, I am perfectly fine with
recalling
the data and the cards. It's cheaper for me.
Joel Goergen
|