No, actually. His point (one) was that startups don't do any good if they manufacture offshore. That exports jobs and technology both, since the technology gets refined over time (with experience).
He thinks we're selling our seed corn. Could be.
-- Cheers, James Arthur
.
actually
maximum=20
if
statistical
=20
they=20
Tabby confuses testing every lot with testing every device.
?-(
diodes
How do you get that? It is a different machine, with a different = capacity in units per hour.
I cannot say that it does or does not support your position.
It was a really long time ago. Decent clean rooms even back then. =46airchild it was. Maybe it was part of an IC fab though.
actually=20
maximum but=20
if=20
statistical=20
would=20
issues=20
diodes
needed
If you have a failure rate of 0.1 % on simple parts like diodes and transistors, you have such process control problems that is time for a major line stop and overhaul to find/fix the problem(s). =20 If you are talking about things like x86 processors 0.1 % failure rate is probably ok. Functional parts are speed graded based on wafer tests. Or so says my ex-Intel friends.
?-)
actually=20
maximum but=20
if=20
statistical=20
the
20=46or the HS 601 program at that time. Not memories, maybe processors = (but i think maybe not), not SSI, not discretes, not passives. Statistical = Lot Accptance Testing on everything; and they did test the living s*it out of those parts. I didn't see any vacuum devices though.
?~/
actually=20
maximum but=20
source
production
space),
if=20
statistical=20
device,
sheets
the
20while.
The Paper trail.
josephkk
actually
maximum but
source
production
million
space),
at
realize if
statistical
device,
sheets
the
20while.
They eliminated the particular test that would have caught that = particular flaw in the name of Cost-Cutting for Sen. Proxmire.
I claim you have misread it. The described testing is for new product qualification/characterization. Not for production testing.
Unfortunately, this does not surprise me. I am glad i don't remember all the testing horror stories and mega-oopses i saw.
transistors.
still).
In compensation you can have a problem with counterfeit whatever you are buying.
?~)
And there are physical reasons for that as well.
: .
...
yum
if
ical
they
I think you'll find this has been well addressed already.
issues
Well, maybe 10% would justify a serious process overhaul.
Over 50% failure rate is common on high-end digital ICs. I've heard of
90% in early production. It's not easy to get a billion transistors right, fabbing features a fifth of a wavelength wide. The fundamental idea of semiconductor economics is to fab lots of parts on a wafer and test, test to find the good ones.John
but
OK, if you don't test the parts, what does the paper say?
John
but
Wait! JAX-TX parts are burned in. So what's the benefit of burning in and temperature cycling parts if they are never tested? How does that improve reliability?
John
but
The level of error on the P-E mirror could have been caught by the kinds of tests that amateur home mirror grinders use. Or they could have tried to make an image of a real or simulated star. Kodak did it right. The problem wasn't cost; they used a very expensive state-of-the-art laser test system, but they used it wrong. The problem was arrogance.
Read the book.
John
still).
One doesn't get counterfeits from a major, authorized distributor. Maxim parts are mostly sole-source by design, and they aren't very good about ensuring a steady supply, so people are forced to go to brokers.
So, design in multi-sourced parts and buy from authorized disties.
John
but
issues
50% was pretty good on some products I worked on. Early on in a process node, productivity was often measured in wafers per good die. At 10% they'd have a sellable product.
It's helpful to read datasheets. Something like a cheap opamp or TinyLogic gate will usually have tiny footnotes that say that some parameters (like pin capacitances) are *not* tested on every part. And some parameters, like leakage currents, are spec'd three or even six orders of magnitude above typical values, because that saves milliseconds on the test machines.
The Agilent 34401A DVM design is interesting. It's a multi-slope ADC that uses cheap HC4066-type analog switches everywhere. If any of those parts has a thousandth of the datasheet leakage, it wouldn't work.
John
but
Statistics, obviously.
Have something to add? Share your thoughts — no account required.
Ask the community — no account required