[sdiy] Pet Peeves

[Florian Teply]

Am Fri, 4 May 2018 21:38:27 -0700
schrieb rsdio at audiobanshee.com:

> We clearly live in an age where companies are designing chips without
> the level of testing or guarantees that we used to rely upon.
> 
There are a multitude of things now playing a role in that game that
more or less were unthinkable twenty, thirty years ago:
* Silicon manufacturing is a highly specialized business. Has been like
  that for decades. But what has changed is access to manufacturing.
  Back then, everything from chip idea up to sales and customer
  service was done within the same company. And quite often it was the
  same engineer or at least the same team who took the chip all the way
  from inception to production including documentation. Nowadays,
  there are often several companies involved, from the design house
  which has no intimate knowledge of the actual manufacturing
  process, through fab staff which doesn't have the faintest idea of
  what Chip X is supposed to do, in some cases several test houses
  running each just parts of the test such that nobody can actually
  infer aynthing on the operation of the chip just from test results
  from one test house alone, to the sales company which has contact to
  the customers and somehow comes up with what might be considered a
  datasheet by non-tech managers. That way only the most often asked
  parameters actually make it into the datasheet. And companies have to
  rely on the help of field application engineers covering all the
  omissions.
* Testing is expensive. I just had the opportunity to learn this myself
  as I am characterizing low frequency noise on just a simple MOS
  transistor. One single meaningful noise measurement easily takes a
  minute or two. Add settling time of the necesary filters on the same
  order and you'll get to about five minutes per device and bias point.
  Eventually I'd like to put reasonable numbers into the simulation
  model for the device and I'll have to wait for the yearly maintenance
  period to actually do the measurement as five minutes times two dozen
  bias points times thirty different device geometries times a good
  dozen sites per wafer times at least three wafers from three lots
  evaluates to a LOOOOOONG time. So if one can justify not measuring
  something by not putting it in the datasheet in the first place, that
  can save quite a lot of time-to-market and money. What's not spec'ed
  doesn't need to be tested.
* Simulation models do not come for free: One has to understand the
  different models including their limitations pretty well in order to
  choose the right simplifications to get to a model that's actually
  useful. To do that properly one would first have to understand the
  device itself, and second also have some experience in circuit design
  and simulation to get an idea of the various applications the device
  might see. It's something like a permanent rehearsal as a new
  apllication idea might depend on device behaviour that hasn't been
  considered important (and therefore not modeled at all or not
  properly) before. It also needs a good deal of understanding of
  modeling on the customers side to notice that this or that special
  case is not included in the model in the first place. I do know of
  companies which generally create custom simulation models of
  commercial devices for their internal use. These models then often
  come with a different (not necessarily smaller) set of shortcomings
  which is tailored to their needs compared to the official
  manufacturers model.

I still don't like the fact that datasheets are getting more and more
sparse to the extreme that they can be applied to a whole family of
devices without change. I'd even prefer to not have a second source for
parts in exchange for good documentation and testing.

In the end this all is driven by the expectation that the monetary
implications of moore's law are expected to apply to analog chips as
well, even though these don't scale well. 

> I mean, chips are such a commodity that people are stealing duds and
> reprinting them with expensive part numbers on them just to scam the
> market. Is it a stretch to say that putting out a chip without a
> complete data sheet is somewhere on the same spectrum, a little
> closer to the fakes?
> 
It's definitely not a stretch to say that. Into the same direction goes
putting out incomplete datasheets with the proposition to get the real
thing under NDA for IP protection reasons. If something is sufficiently
secret to justify that it should not be in a public datasheet, it
probably isn't a good idea to put it publicly on sale in the first
place. 

So much for my part of the rant....
Florian Teply