Re: PCI card current loads

[Bruce Hanahan <bhanahan@ccmail.crosscomm.com>]

     Please see my reply interleaved below.


______________________________ Reply Separator _________________________________
Subject: PCI card current loads
Author:  pci-sig-request@znyx.com at inet
Date:    10/17/96 8:54 AM


We are in the process of designing support for multiple PCI busses in a 
high end UNIX/NT system.  Our power supply guru has the following 
questions that he was unable to answer from reading the PCI spec. Anyone 
have the answers?
     
     
I need to know two things about the PCI cards we will be using in the 
next platform. Ideally, I would like to know worst-case values, but 
realize I may have to settle for typical numbers.
     
1) What is their dynamic load current characteristic, at the card level?
     
For example, assume a 5V card drawing an average power of 20W. This could 
occur with a DC current of 4A, which would happen if all logic activity 
was perfectly filtered out by on-card capacitors. More likely, there would 
be some lesser DC current with an AC component riding on top. The question 
is: How big is this dynamic component, and what frequency is it at? I do 
not care about individual gates and signals, but rather the combined 
result of everything on the card.
     
Is there a different answer for different cards (3.3V only, 5V only, 
Universal)?
     
     ********************************************************************
     1.) This is not covered in the PCI spec because it has nothing to do 
     with PCI.  This is a question of how the board is designed.  If you 
     were to design a two layer board, for example, you would have a much 
     worse AC component than you would with a four layer board because the 
     power supply and ground busses on the two layer card would have much 
     higher inductance than the power and ground planes on the four layer 
     board.  
     
     The AC component of supply current is a function of the ICs on the 
     board and their switching frequency.  There is usually some section of 
     the IC spec that gives dynamic current guidelines based on switching 
     frequency and how many outputs are driving at a given time and in what 
     direction.  In my experience you have to sort of interpolate your 
     dynamic current based on the guidelines in the IC spec.  It is 
     extremely difficult to get an exact number because although the part 
     may have a master clock input the outputs usually switch in a random 
     manner, which makes the output switching frequency much lower than the 
     master clock frequency.  You should be able to ballpark it well 
     enough, though, for most applications.
     
     However, the dynamic supply current would not be a factor at all in a 
     perfect world with perfect board bypassing, except for having to 
     source the dynamic current from the supply to recharge that perfect 
     bypassing.  Since we don't live in a perfect world (at least I know I 
     don't!) the amount of AC ripple present on the DC supply voltage is 
     strictly a function of the AC current and how well the board bypassing 
     is designed.  Again, this has nothing to do with PCI or the PCI spec, 
     but is totally dependent on how well the board is designed.  The board 
     designers have to answer that question, not the spec.
     
     I highly recommend buying a copy of Howard Johnson's "High-Speed 
     Digital Design (A Handbook of Black Magic)" from Prentice-Hall.  This 
     is an excellent reference on not only bypassing and board design, but 
     on all high speed digital issues, like crosstalk, transmission lines, 
     terminations and the like.  I bought my copy on a trip to San Jose at 
     the Computer Literacy bookstore (sorry, this isn't an ad, just a 
     reference).  
     ********************************************************************
     
     
2) What is the amount of bus capacitance, at the card level?
     
This will directly bear on how well the card filters out its dynamic 
switching currents, but I also care about how much capacitance the power 
system must charge during turn-on. Again, is there a different answer for 
different cards (3.3V only, 5V only, Universal)?
     
     ********************************************************************
     2.) If I understand this question correctly then you are looking for 
     the capacitance of the PCI backplane.  Again, this has to do with your 
     backplane design, including the capacitance of the pc runs, the 
     capacitance of the connectors (which should be spec'd by the 
     manufacturer) and the sum total of the capacitance of all the PCI 
     drivers on the bus (again, should be spec'd by the manufacturers of 
     the ICs driving the bus).  The runs on the bus should be electrically 
     short, so the capacitance can be modeled as a lumped capacitance (just 
     add them all together).  If the lines are not electrically short then 
     the capacitance must be modeled as distributed.  This distributed 
     capacitance will lower the characteristic impedence of your pc runs, 
     which will affect your PCI drivers.  The PCI spec *does* cover the 
     amount of min AC impedence the drivers must handle (I believe it's 
     around 30 or 32 ohms) and you calculate this number and not violate 
     it.  Again, reference the Johnson book for a good tutorial on this 
     topic.
     ********************************************************************
     
 --------
     
The PCI spec is very vague and states only that "All power planes must be 
decoupled to ground in such a manner as to provide for reasonable 
management of switching currents to which the plane and its supply path 
are subjected." Refer to section 4.3.4.3.
     
Thanks in advance for any help you can provide.
     
Kirk Soldner - Power Design

     ********************************************************************
     Sorry if the replies were a little too verbose. I hope that they helped.  
     Best of luck, these kinds of issues can be a real pain in the neck.  
     
     Bruce Hanahan
     bhanahan@crosscomm.com
¦ˆw