[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: M66EN and PRSNT signals (re-send)



Signal integrity 101 ...

To understand why pins such as M66EN and PRSNTn# need decoupling, you have to "follow the electrons," or the paths the current takes.  We may live in a voltage world, but current and voltage always accompany each other (except for static electricity or superconductors), and often the current is more important.

At DC, currents spread out and it may not make much difference where it flows.  If you provide enough power and ground pins with margin so you don't "smoke" them, and so the DC voltage drop is OK, you could group the pins any way you want.  But at AC, which means we are talking about the initial transient response when signals switch, it's a different story.

Let's look at the PCI connector.  When a signal pin switches from (say) low to high, driven by the plug-in card, there's a current that goes through that signal pin to accompany the voltage transient.  For the sake of illustration, let's say that a signal was quiescent at 0V, then switches to 5V, and that this card is in the middle of a long-ish PCI bus with motherboard traces in both directions from the connector, and that the traces have a 70 ohm "characteristic impedance" (due to their trace width and proximity to ground planes).

When the signal was quiescent at 0V, its current was 0A.  When it goes to 5V, the signal current quickly becomes 5V / (70ohm / 2) = 143mA.  (This is somewhat exaggerated, because in reality we wouldn't see this much signal current, nor as much voltage swing all at once, on a typical PCI bus signal.)

Add 30 more pins that switch at the same time, and you have a lot of current being switched.

This current is transient; it lasts until the signal reflects off the ends of the PCI bus and bounces around a bit.  But for the moment, we have that much current flowing through the signal pin.

Currents always form a loop, a complete path.  When current goes out this way, an equal amount must come back that way.

So there is a return current somewhere.

Return currents are everywhere.  Any time a signal goes down a trace, there is a return current that mirrors it, preferably in the ground plane under/around the trace.  At AC (meaning, we look at transients, when signals switch), the return current closely hugs the signal current.  If we don't provide a nearby path for the return current, we get evils like crosstalk, impedance mismatches, even radiation.

But there's no ground plane through the connector.  Even if the card has four or more layers and you continue the ground plane under the finger pins, it still stops there.  The only path for ground return currents, is through other connector pins.

The pinning on the PCI connector was such that, over most of the connector, every third pin is a power or ground pin.  Thus, each signal pin sees (at least) one adjacent power or ground pin, which is *guaranteed* to be connected to planes or well bypassed on both boards.

So when signal pin X suddenly has 143mA going through it to the motherboard, there's a place for the mirror 143mA to come back onto the card; namely, the adjacent pin.  (In reality, other nearby pins also share in this return path, but the adjacent power or ground pin is crucial.  Otherwise, crosstalk can become intolerable.  To some degree, the overall density of power, ground, and bypassed pins like PRSNTn# and M66EN, plays a part too.)

If you don't bypass the PRSNTn# pins, then INTD# and a Reserved pin (which might be used later) would see no adjacent return path.  Oops.

M66EN was a ground pin, before 66MHz PCI.  When it got liberated to become M66EN, AD[10] potentially lost its return path.  To restore the return path, M66EN gets bypassed too.

They aren't really so unique.  You can think of them like all the power pins, which are required to be connected and bypassed as well.

(Which is why I cringe whenever I see a PCI card that's missing finger pins!  It may work in some systems, but the margins are reduced and it could become responsible for those occasional crashes and hangs, or corrupt data.)

Sorry about the long-winded response, but I hope this helps answer your question.

Regards,
Andy