[gentoo-amd64] Re: Re: partly OT: nvidia 6600GT

[Duncan]

Roy Wright posted <4401999F.9080507@×××××.com>, excerpted below,  on Sun,
26 Feb 2006 06:05:51 -0600:

> Peter Humphrey wrote:
>
>>Which reminds me of a problem I had a few years ago which was solved by
>>moving a card to another slot. When clutching at straws ...
>>
> I had a similar problem last year, turned out the 6800GT wasn't all the
> way into the connector.  When I would tighten the hold down screw, the
> opposite end would raise just a little (~1/16").  Removed and reinstalled
> the board three times before I caught it...

It's observed that the following discussion probably isn't new to many,
but perhaps some haven't thought about all the complex cross-effects in
this way before...

Actually, if you read the troubleshooting info with quite a number of
items, both reseating and trans-slotting are officially recommended
troubleshooting methods.

Over months or years of on/off cycling and other hot/cold cycles, not only
is it possible for cards (and connectors) to work loose, but an oxidation
layer may form preventing a decent connection.  Again, at the frequencies
many computer components operate at, the problem may not be simply lack of
a good DC electronic connection, but the fact that the changing
conditions takes a connection or set of connections out of capacitance
spec, sometimes causing an entirely separate resonance at an unintended
frequency.  Capacitance of course allows current to jump to parallel but
not directly connected circuit traces, and there are often lots of those
on a circuit board.  The circuit of a specified frequency intended to be
completed over one set of physically connected leads may end up entirely
transmutated as a parallel circuit develops resonating frequency
capacitance, changing the character and details of the transmitted data
entirely and wrecking havoc.

Note that as memory speeds have increased, mobo makers have had to go to
increasing lengths to ensure such interference and cross-echo patterns
don't develop.  Properly terminated data circuits aren't just a SCSI
thing, and many a board design has been delayed or even ultimately junked
as it simply couldn't handle the memory and other signal timings, and keep
trans-circuit resonance and feedback, to within acceptable tolerance
levels.

Anyway...  something as simple as reseating a card, due to the friction of
the connector spring pins against the exposed board traces, will often
scratch away that oxidation and return the various circuits to designed
operating parameters in both DC resistance and RF capacitance, within and
across circuits.  That's often all that's needed, and the reason just
taking something apart, finding nothing visibly wrong, giving up and
putting it back together "without changing anything", actually allows some
things to work again.

As for trans-slotting, it's less common in modern boards, but it was quite
common a few years ago for only 1-2 slots out of 3-5 to be fully
"pci-master" capable.  Anything using heavy data transfers could often
make use of pci-master, offloading work from the CPU under certain
conditions, altho the most common use was data transfer between harddrives
-- if both were in PCI-master slots, or one was and the other was onboard,
the chipset could often handle data transfer between them without direct
CPU intervention, once the transfer was initially setup.

Additionally, the PCI spec has four interrupt channels, normally staggered
so a different one will be in each slot of the first four -- a fifth slot
shares placement with the fourth, in most cases, and will work best  when
purposed to something (like a video card) not normally requiring
interrupts.  This is because while all four interrupts are available and
exposed to all slots, most cards only use one primary interrupt pin, in a
standardized location, altho some will allow use of a second for
flexibility and many multifunction cards use two or more if all functions
are enabled.  Most new mobos come with a diagram of their PCI interrupt
layout per slot, and specifically advise trying a different slot if a card
doesn't work as expected, thereby causing it to utilize a different one of
the four PCI interrupts.

Finally, there are intercard proximity factors as well.  Of course,
there's pure physical space requirements -- some cards simply don't fit
next to each other or for sound cards and the like, plus PATA/SATA drives
and even interface cards with detached plugs, attached cabling simply
won't reach to all possible slots.  Heat and ventilation are another major
factor.  However, the same cross-circuit inductive and capacitive effects
outlined above can play a factor here, as well.  This can be as simple and
easily recognized as a buzz on your sound card when it's too close to
another specific card, or as difficult and unintuitive to recognize as 
memory issues or disk storage errors.  One particular thing that we've not
had to worry about for a few years (due to dedicated AGP slots) that we'll
have to start worrying about again as PCI-Express becomes more common is
video card placement, as those have all sorts of RF things going on, with
both high data rate memory access, and high frequency video signalling. 
Get those signals cross-interfering with something like memory/CPU
channels or CPU/CPU interconnects (AMD's hypertransport, and Via has a
similar technology for northbridge/southbridge interconnects, among
others), or an add-on SATA/PATA/SCSI interface card, and there could be
/major/ data integrity issues to worry about!

It should be quite obvious by now that mobo designers certainly don't have
it easy, these days!  Sure, they deal with lower frequencies than CPUs and
graphics chips deal with, and don't have the intricate logic and heat
dissipation layout issues that those guys do.  However, the frequencies
they do deal with are quite high enough to cause **MAJOR** cross-circuit
interference design issues, particularly over the trace length of
centimetres they deal with, as opposed to the nanometre/micrometre scales
a chip designer deals with.  As I said, it's not unusual at all for a mobo
design to simply not be able to handle the necessary timing demands placed
upon it, and either be limited to the low end of an offered range (those
mobos that are certified for PC2700 memory or less, because they simply
can't handle PC3200), or end up entirely rejected due to stability issues.

Under this sort of tight timing requirements, it's actually a wonder we
don't have more cards that simply won't work right in certain boards,
while they work just fine in others.  That alone is tribute in itself to
the wonders of human engineering genius, both in design, and in the
ability to actually implement said design in an assembly-line environment
to within the required tolerances, without costing us several thousand
dollars for the mother-board alone!

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman in
http://www.linuxdevcenter.com/pub/a/linux/2004/12/22/rms_interview.html


-- 
gentoo-amd64@g.o mailing list