| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Hi Pat, I wasn’t taking about 90%. I was
taking about 99.99+%. If the worst-worst case was at 10 sigma, would that be
over design? Can you even calculate 10 sigma on a hand calculator? J I use all sorts of distributions in my
multi variable assessments. It turns out that, when I have 5 variables, the impact
of distribution choice is small. When I hit 10 or more variables it is irrelevant.
It’s sort of the “Central Limit Theorem”. However I do understand what you a saying.
There statically variations and blocking conditions. Our challenges it to
determine which are blocking conditions. Is drive voltage blocking or stochastic?
Perhaps we can get consensus on this at April
19th meeting if you like. ... Rich From: Pat Thaler [mailto:pthaler@BROADCOM.COM] Rich, One can't assume that
implementations are done such that the limits in the spec are at the 3 sigma
points with the mean about the middle. Implementations won't necessarily be
built to the average of the spec. For instance, I've seen cases with past
standards of interoperability problems where one vendor chose to put their
transmit level close to the minimum with tight standard deviation (trying to
optimize for emmissions, power consumption, etc.) and another vendor put their
input sensitivity near the maximum. Another example is that sometimes one
or more of the specs is particularly difficult to achieve and designs have an
average value that is very close to the limit. A vendor may even take a
yield hit and sort out those chips that are over the limit so occurance at
or near the limit value happens much more often than 3 sigma would
predict. In the past standards I've worked
on, we have usually assumed that implementations could be running at the limits
because we don't have a basis for assuming a distribution. Where we have
applied distribution, it has been to channel characteristics for which we had
some basis for believing there would be a distribution - e.g. not all channels
will have the worst crosstalk situation. Also, a typical backplane system
will have a lot of links. If 90% of them work, the system still has a problem.
We have to do a good deal better than that. Pat From: Mellitz, Richard
[mailto:richard.mellitz@INTEL.COM] Ya know… I just did a statistical
analysis of the probability of units failing the return we just voted on. Under
some assumptions I made, I came up with 2.5 units per 1000 would fail RL and
still pass. I heard folks thought this might be something like 90%. The design
question is what quality level is acceptable. I understand this from a
business perceptive because I can relate it to cost. I don’t know
how to apply this to standard work. Maybe it does mean all limits at the worst
case must work regardless of the likelihood. Maybe this is a Pandora’s
Box too. I think our cost is not dollars but delay producing a workable
standard. The 80% Joel was talking about was design
engineers. This is not the statistics of a design’s quality. If I +/-3
sigma all our limits in the spec, I think we are more like in the 99.9+%
quality range right now. The task at hand was to determine if the
informative channel spec sufficiently predicted confidence related to the EIT
receiver test. That why we used the term “confidence” and not
“limit.” Remember that is why we chose the channel to be
informative. We showed we couldn’t constrain all 3 (tx, channel, rx) and
create a reasonable and marketable solution. So in light of that I believe we
should constrain the analysis to reasonable. Maybe we should do it both ways
and discuss what is reasonable at April 19 meeting. … Rich |