http://www.adexelec.com/faq.htm#max_length
"What are the maximum cable lengths on flexible extenders?
PCI 32 BIT 33 MHZ 5 INCHES"
Some examples of their extender products...
http://www.adexelec.com/pci32.htm
The PCI bus is a reflected wave bus, without terminations on it.
The signals go down the bus, bounce off the end, arrive back at
the origin, and at that point, everybody sees a full amplitude
signal. Then the clock can rise and take a data sample. The signals
continue to bounce around, so the situation is pretty "chaotic". But
PCI is a very cheap bus, and easy for people to build to, which
is why it continues to be so popular.
http://en.wikipedia.org/wiki/Reflected-wave_switching
The original bus configuration, driver impedance characteristics and
the like, were selected after Intel did something like a thousand
hours of analog simulation. The simulation environment takes into
account the possible bus population cases (few slots used, lots of
slots used), and presumably some limit was set as to how many slots
they wanted to support.
In a project we were working on, where the conditions were more
controlled than your average PC, we were able to make the bus longer.
But to do what we did, we used analog simulation to see what was
possible, taking all the extension segments, connectors, lack of
stubs (flyby routing) into account. Then the bus could run longer.
Our bus had no provision to remove "cards", and all the loads
on the PCI bus were permanently soldered in place. There were
fewer connectors in the paths as well.
Generally speaking, there isn't a closed form expression that will
tell you how long the bus can be, which is why simulation is
used extensively, to predict what configurations will work.
Even though the company listed above, makes a variety of
adapters, they probably don't make any statements about how well
they work, or exactly how carried away you can get. So if
you were to continue with that proposed solution, you may
want a fallback position. The fallback position, is to reduce
the PCI bus clock. But not all PCI devices are designed to work
all the way down to zero hertz, so there can be limits as to
how low the clock can be dropped. And the motherboard clock generator
may not have a useful setting, or a means of easily accessing it.
I no longer have access to good simulation tools, and there
aren't good enough models in the various versions of free SPICE
programs, to make predictions with. So even if I was curious
today, I cannot satisfy my curiosity and answer your question
with a simulation. The tool I used to use, cost something like
$35K per year for a license.
If you want a commercial expansion solution, have a look at this company.
http://www.magma.com/products/pci/4PCI/index.html
What they do, is use a PCI bridge from the motherboard, to
their expansion chassis. That allows the bus to be extended
without breaking any rules. They have several host adapters,
so you can even connect one of their expansion boxes to a
laptop that has an Expresscard slot. The host adapter,
decouples the host bus from the extension cable. That allows
the bus to be extended, without breaking any electrical design
rules. The price paid, is higher first cycle latency on the
bus, which is only an issue if the cycle mix consists mostly
of single cycle transactions. As long as hardware does
lots of DMA, done in bursts, you hardly notice the
impact of the bridge card. It is only during configuration
operations for the card in the expansion chassis (which may or
may not be occurring all the time), that you'd see the less
desirable (slower) single cycle type transactions.
You can see the pricing here is pretty predatory, with a single
slot box costing about $1000. There are two expansion "box" options,
and six different host adapter types (to extend from the host
system), shown on this page. The extension cable they offer,
is 1 meter long, so the expansion chassis could sit on top of
the computer perhaps. Presumably, the expansion chassis has its
own power supply as well.
http://www.magma.com/products/pci/1PCI/prices.html
So you can (and undoubtedly will) continue on with
your project. Just be prepared for surprises. I'll
give one tiny hint - the best way to extend, is to
extend from the last slot on the bus, as at least
that won't leave a humongous stub in the middle
of the bus. If you "fork" the bus, like in the
following example, I can't even guess how messed
up the signals will be.
Southbridge ----X--X--X--X--X <--- now a big stub (bad)
|
+------ long extension ---- PCI card
I'd prefer to try it this way, if I was going to attempt it.
Southbridge ----X--X--X--X--X (better)
|
+------ long extension ---- PCI card
It might also help, if fewer cards were plugged into the
main bus. That might improve the time of flight (
reduced loaded delay), allowing the bus to settle and
meet Tsetup.
As for the impedance of the extension, that could cause
a reflection too. The various kinds of cabling solutions
have different target impedances. Ribbon cable might be
up around 110 ohms or so. The Kapton film that Adexelec
shows, flexes in only one dimension. But with the film, you
can tune the electrical performance as you want. Short
Kapton assemblies can cost a couple hundred in small
quantities, so the price will be a deterrent. Ribbon
cable on the other hand, you get what they give you.
On ribbon, you can fool around, by alternating gnd-signal-gnd,
to change the conditions a bit. Or switch from the normal
insulator, to something like Teflon (drop to 90 ohms perhaps?).
Being a pragmatic guy, I'd probably take my Sawsall, and just
cut a hole in the PC side cover, so the full sized card
can just pop its head out
At least that is solely a
mechanical project, and not electrical.
http://www.curtislumber.ca/products/graphics/tools-sawsall.jpg
Good luck - you'll need it,
Paul
