[time-nuts] Reliability of atomic clocks

jimlux jimlux at earthlink.net
Mon Mar 28 10:51:44 EDT 2016


On 3/27/16 8:20 PM, Mark Sims wrote:
>> Mil-spec parts would be somewhat more reliable than commercial
>> parts.
> Actually,  that is seldom true.  The main difference between mil-spec
> parts and commercial parts tends to be in the post-packaging device
> testing (e.g.. extended temperature / voltage range).  They usually
> have the same guts inside the package.    I friend of mine was a test
> engineer at a major semiconductor manufacturer and he said their
> commercial parts were actually more reliable than the mil-spec ones.
> The commercial parts were built in vastly greater quantities and
> their production flow / experience enabled them to continually fine
> tune the process and testing.  The production / test regimen of
> mil-spec parts tended to be locked-down by inflexible specs and
> procedures.  Changing anything required re-qualification of the
> entire process.
>


Back in the day (e.g. 70s) there was a lot more process variability, so 
the whole screening process involved in selling 883B parts (54LS vs 
74LS, for instance) made for lower infant mortality, and smaller 
performance variations (e.g. if it had to meet performance specs from 
-40 to +85, the performance at 25 would likely have smaller variance 
than for a part that wasn't necessarily tested at any temperature)

There was also the thing of the perceived "higher reliability" of the 
packaging: ceramic packages with leak tests vs plastic encapsulation. 
(for space applications, I'm always intrigued by folks being concerned 
about leak test performance.. the parts going to operate in a vacuum, 
after all; what the leak test performance is really about is as a 
manufacturing process quality assurance, not because you actually care 
about leakage).

This concern with consistency is where Source Control Drawings and the 
rigorous process management comes from for high rel parts.  I'm not sure 
that this is such an issue today.

These days, processes are MUCH more consistent: if you're making deep 
submicron geometries, you've got to have your process under control to 
get acceptable yield.

As noted above, it's probably the same die in the package whether it's 
flipchip/chip-on-board/plastic/ceramic flatpack.

However there is a whole new issue of "fabless mfrs" - your design's 
performance might depend on some non-datasheet aspect of the part 
(radiation tolerance in particular) and if the new parts come from a 
different fab, or if the fab changes their process, that non-datasheet 
performance might change.

In my business of space electronics, we build tiny volumes, and tend to 
make the same design for years: this is not a big money maker for any 
chip supplier.  And when it comes to parts selection, it's easier, come 
design review time, to just use whatever venerable part you used before 
than to justify selecting a new part. This gets us into trouble all the 
time, when that trusty part was end-of-lifed 10 years ago, and we've 
been building units from new old stock we bought 20 years ago, and now 
we're out of parts.  The software defined radios we fly to Mars use 
Xilinx Virtex 2 FPGAs, which end-of-lifed a few years ago, yet that's 
what's likely to fly in 2020 and later.  Not only do you have to 
scrounge parts, but getting and running development tools is non-trivial.


Given the huge market these days, and the tiny, tiny market for 
MIL/Class-S high rel, most mfrs aren't interested in providing detailed 
information on the fab and processes (if for no other reason than it's 
the proprietary secret sauce). Nor are they particularly interested in 
following some process approved 30 years ago for a part on which they 
won't make any money.

There are firms like Rochester Electronics (leader in the trailing edge) 
who fill this niche, they stockpile obsolete parts, get manufacturing 
artifacts (masks, etc.).  but they aren't cheap.

These days, the automotive industry really drives high rel extreme 
environments: under-hood engine control units have to be very reliable: 
a failure rate of 0.01% would be crippling for most auto manufacturers. 
  Some consumer electronics also has to be fairly reliable, at least for 
infant mortality (warranty returns cost money)




More information about the time-nuts mailing list