Controlled impedance for PCIe on carrier board


We’re using the Jetson TX1 module and have designed our carrier board to also accept PCIe full mini cards. We’re using the 2.5GT/s rate. While we are able to achieve a successful PCIe link between the TX1 and the add in card, we also observe significant EMI interference which is desensing a nearby GPS receiver (~1.575GHz) when the PCIe link data lanes are signalling. We’ve narrowed down the issue sufficiently to suspect a relationship with the PCIe data lanes themselves, but not necessarily a root cause.

After reviewing the TX1 design guide, I noticed that the recommended differential impedance for the PCIe link was not followed by our layout engineer. It was routed at 95 ohm differential, as opposed to the recommended 85 ohm. Intrapair matching was exact and all other design rules were followed (except voiding of connector pads and ideal GND via placement. GND return vias are on the same grid as the Jetson connector pins so aren’t ideal.)

What is the 85 ohm recommendation based on? Is the Jetson TX1 module itself routed at 85 ohm? Is the source impedance of the transmitter in the TX1 chip 85ohm? If not, what are the differential impedances of these elements?

Thanks very much,

I can’t answer the impedance question, but this is related…

I’m not sure how you’d check, but default on a dev board for PCIe is to enable spread spectrum (SS), which has lowering of EMI as its purpose (if you use the stock kernel config then you will have SS enabled). If you’ve disabled SS, then EMI would probably increase. Note that although a Jetson has SS enabled by default that most desktop computers only offer SS as an option and do not normally enable SS. If you go to tune impedance you probably want to do so with SS disabled, and when done testing and if you still need reduction of EMI, then you’d re-enable SS.

Hi Mike,

The impedance is to match that of module routing, there are +/-15% tolerance for it, 95 ohm is in the range.

Generally if the GPS lines are on the same layer and nearby, a shielding GND line to isolate from PCIe lines is necessary, if on the different layers, the covering GND plane is necessary.

@Linuxdev, thanks for the tip. I’ll let our software guys know, that will be useful for debug efforts! They’re also looking at the registers for adjusting amplitude, skew, etc, at the PHY level.
@Trumany, thanks for the info. Our GPS receiver is on a different board. The coupling is actually radiated in our case. We can use a standalone GPS receiver mounted about 10cm from the TX1 and see that it, too, suffers the same magnitude of desense. So we’re trying to assess if there’s a significant enough impedance discontinuity to cause common mode conversion and thus radiation.