I have ALWAYS distrusted simulation and computer modeling. And I used to
teach the stuff. GIGO.

"Neither the voice of authority nor the weight of reason and argument are
as significant as experiment, for thence comes quiet to the mind."
R. Bacon

-John

===============

Bob Camp wrote:

there are also designs that depend on "non-data-sheet" performance of
particular devices.  There's a very low noise, very low leakage fet
popular in charge amplifiers.  It has a JEDEC 2N number (which I can't
remember off hand), but only the ones from one particular company (in
England) actually work in the circuits, and even then, there's some hand
selection involved.

FWIW, IMO any engineer who uses undocumented or uncontrolled parameters or
instructions in a production design is a fool.

If you are that silly, you must fully specify the selection criteria.

-John

===============

(in production yes I agree)
In research things are different.
You wouldn't mind to make an selection of fets or else
to obtain the very top specs of a certain unique instrument
design as the real final product are the results you
may obtain with that instrument and not at all it's design...
So many nice stories about gear that worked only with a certain set
of parts way off the manufacturers expressed data...
one day I'll drop a few here just for amusement.

Luis Cupido
ct1dmk.

J. Forster wrote:

Simulation has some value in determining things like allowable component
tolerances and Worst Case analysis, but those are really production
engineering rather than design.

As to working at brassboard but not in production, it is prudent to check
that your parts are within the production part specs.

John

===========

J. Forster wrote:

Or, has their back against the wall and can't do it any other way.

How is this any different than using trimpots or hand select?

For years, folks have hand selected matched pairs of devices, since the
circuit requires tighter tolerances than the mfr guarantees.

Many, many RF designs have "select at test" pads to set levels or tuning
stubs depending  on what the actual gain or impedance properties of the
active devices are, or for trimming temperature dependencies.

Would you say that the engineer is a fool for not just specifying
tighter tolerances.. the tighter tolerances may not be available from
the mfr (who has to respond to many customers, most of which will be
happy with the standard performance).  It's sort of a tradeoff.. do you
go to the mfr and say, I need a better grade of part, or do you buy the
run-of-the-mill part, and sort them.

You might decide to do the latter for competitive reasons, e.g. rather
than the mfr producing a better grade of part, and potentially selling
it to your competitors too, you keep the "secret sauce" in house.
(Granted you could have the mfr make/select a proprietary part for you..
that's basically changing who does the work, but doesn't change the
underlying design)

Even manufacturers do this, for instance with speed grades on things
like microprocessors.  They don't have enough process control to
guarantee a particular speed, so they make em all, and then sort them.

The other thing is that the selection criteria might not be knowable in
a standalone sense.  That is, you have to put the part into the circuit
and see if it works, rather than measuring some device parameter.  I
would agree that to a certain extent, this implies that you don't really
know how the circuit works, but it might also be that the most cost
effective approach is to use empiricism, rather than analysis.

On 08/14/2010 05:08 PM, J. Forster wrote:

Using non-spec aspects needs documentation, motivation and risc analysis.

I've designed a selection routine for a component which still ticks on
with good statistics. The main problem is that eventually that component
will be on last buy level. Obviously it seems the selection was chosen
on the conservative side, so it works in shipped products regardless of
batch. Trimming of the manual routine has lowered a certain failure mode
of testing.

Any spec should be verified. Published specs needs verification with
real components. Unpublished specs needs consistency testing or even
selection testing on all components. Cost of testing needs to be
understood and risc of low yield in future needs to be understood and
alternative approaches could be put in place before running on flat tires.

At times it may be cheaper and safer to run with more expensive
components which is within spec.

Cheers,
Magnus

I think you missinterpret what I meant. Two examples:

I've seen programmers who use "instructions" that are not part of a uP
instruction set and are undocumented, just to be "clever". If a different
brand of chip, or even a different rev., the chip does something
completely different. These guys should be strung up by their tender
parts.

I've also seen transistors used as avalanche switches (basically a failure
mode). If a different production run has improved normal mode performance,
the avalanche function may vanish.

FWIW,

-John

===============