2
\$\begingroup\$

We have 3 PCIe slots on a board, let's call them A, B and C. On this board, slot A will always be populated with a PCIe device, however, among slots B and C, only one of them will be populated.

There is only one PCIe link on slot A. So, in order to communicate via PCIe with either slot B or slot C, the same PCIe TXP/RXP/TXN/RXN signal on slot A should be routed to both slot B and slot C.

I don't know if this would be considered a common scenario. Probably not. I haven't been able to find out any layout guidelines covering this particular case.

Any ideas would be appreciated.

\$\endgroup\$
5
\$\begingroup\$

I would go for either an signal or a packet-switch between these 3 components. With a signal-switch you'll have to detect which card is inserted and activate that signal. If you use a packet-switch (like the ones from PLX) you should be able to do it without any detection.

\$\endgroup\$
1
\$\begingroup\$

You can't do. PCIe is point to point, unlike PCI, which is bussed as you are thinking.

\$\endgroup\$
4
  • \$\begingroup\$ it still is point to point in my configuration. I'm saying that (ATXP,ATXN) are connected both to (BRXP,BRXN) and (CRXP,CRXN), and, (ARXP,ARXN) connects both to (BTXP,BTXN) and (CTXP,CTXN). It is GUARANTEED in the system that only slot B or C will be populated, so in effect A is connected to a single point. It is just that since the traces must go to both connectors, I don't know whether there will be any signal integrity issues. \$\endgroup\$ – SomethingBetter Aug 23 '11 at 18:57
  • \$\begingroup\$ I stand on my previous statement. There is termination in the destination card, so if you are in the middle slot you will have a T going to the other slot. \$\endgroup\$ – Brian Carlton Aug 23 '11 at 19:56
  • 1
    \$\begingroup\$ @SomethingBetter - Are you communicating from slot A to one of slot B or C, or communicating from a main processor on the motherboard to slot A and one of slot B or C? \$\endgroup\$ – Kevin Vermeer Aug 24 '11 at 4:06
  • 1
    \$\begingroup\$ @Kevin: From Slot A to one of slot B or C. There is no processor in this PCB. It is a backplane. \$\endgroup\$ – SomethingBetter Aug 24 '11 at 9:22
0
\$\begingroup\$

The PCIe spec states that small (100nf) capacitors should be installed on the transmit differential pair; This gives you a really easy solution, at least for one direction.

For the other direction, I would just use two zero ohm resistors. Follow the same guidelines that you would use when selectively populating RF sockets. Use a 'T' intersection and the cap or resistor bridges the gap between either the left side and the middle, or the right side and the middle. There will be some loss/degradation but it should be well within acceptable limits for PCIe.

\$\endgroup\$
4
  • \$\begingroup\$ Thanks for the response. In our case, zero ohming is not possible. The end user of the system will interchange the cards. He can not be soldering/desoldering resistors. Also, the PCIe signals are stripline. They never come out to the surface, except of course when they come up to connect to the connector. \$\endgroup\$ – SomethingBetter Aug 23 '11 at 19:01
  • \$\begingroup\$ @Andrew The 100nF caps have to be mounted as close as possible to the transmitter. So they are mounted directly next to the IC which is sending. \$\endgroup\$ – Nico Erfurth Aug 23 '11 at 19:49
  • \$\begingroup\$ it wasn't clear whether these were all soldered components or even user-switchable. In that case I would agree with @masta79; you have to install either a pcie-pcie bridge, or try to do something with signal switching, although at 2.5Gbps that's not going to be easy. \$\endgroup\$ – akohlsmith Aug 24 '11 at 14:32
  • \$\begingroup\$ There is a switch from TI which is capable of handling PCIe gen1 (2.5 Gbit/sec) I'm not sure about gen2. \$\endgroup\$ – Nico Erfurth Aug 24 '11 at 16:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.