Bypass Arrays

We all know about minimizing and reducing, etc. But does anyone out there actually DESIGN their bypassing networks? I sure don't know how to do it!

If I did, perhaps I could design a bypass network that would serve my needs AND the needs of the assembly/manufacturing folks. A few years back, at a SPICE class in San Jose, I heard a fellow from IBM state that he designed his bypassing networks in the frequency domain. Anyone else doing this? I've tried, but without satisfactory results. Any suggestions on scaling the problem? Any ideas on how to model the timing jitter between parts that are simultaneously switching with ~2ns edges? Are there any bypass network designers out there? I'm sure I (at least) could use some real design help on this one. No opinions, intuitions or hand-waving please. I have plenty of my own and they are probably more conservative than need be.

There is a reasonable discussion of how to design a bypass array in Chapter 8 of "High Speed Digital Design", by (this author) Howard W. Johnson and Martin Graham, Prentice Hall, 1993, ISBN 0-13-395724-1.

**Also, I would like to echo the comments of many others.

(1) Adding traces in series with bypass capacitors *decreases* their effectiveness

(2) Adding traces in series with the Vcc/Gnd pins of an IC does *not* help prevent noise from the IC from reaching the Vcc/Gnd

For high-speed digital designs, and also for mixed-signal designs, I place vias as close to the capacitor pads as possible, leading directly to the Vcc/Gnd planes. I also place vias as close to the IC power and ground pins as possible, leading directly to the Vcc/Gnd planes.

Let's look at the example that started this discussion:

                           +----------+
                        ###|          |###
                           |          |
                        ###|          |###
     vias                  |          |
       X=====###========###|          |###
             | |           |          |
       X=====###========###|          |###
            bypass         |          |
           capacitor    ###|          |###
                           +----------+
                               I.C.

Some engineers believe that putting traces in series with the IC power and/or ground pins will reduce the amount of noise that reaches the Vcc/Gnd planes. It doesn't. The reason why has to do with the relative impedances involved, and the problems with ground bounce. Here's a general explanation:

First, consider the chip as a source of noise. The output impedance of the chip (the effective driving point impedance of the chip's Vcc and Gnd pins) is always GREATER than the Vcc/Gnd impedance of a properly designed board (if it's not, then you have a whopping huge amount of Vcc/Gnd noise). This ratio is what controls Vcc/Gnd noise.

Whenever your noise source has a higher impedance than the load, the noise source acts like a "current source". In effect, it pumps a fixed amount of current onto the Vcc/Gnd system.

Putting an impedance, like a little circuit board trace, in series with a current source is futile. It doesn't affect the amount of current delivered to the Vcc/Gnd planes, and therefore doesn't affect the Vcc/Gnd noise.

In order to substantially affect the amount of current going into the Vcc/Gnd planes you would need to insert in series with the chips Vcc/Gnd pins a trace with an impedance *as large*, or larger than the driving point impedance of the chip. But, that's impossible, because an impedance that large would create ridiculous amounts of ground bounce (or Vcc bounce) in the chip.

Now, let's look at the problem in a little more detail.

Imagine that the chip below has eight outputs, and they have all been set to LO for quite some time. At time Tzero, let them all switch HIGH.

If the outputs are heavily loaded, the chip must now source a huge pulse of current. This current flows in to the chip through its VCC pin, and out to the loads through its eight output pins.

From the Vcc plane, this pulse of current must flow along path (A) to the chip. As the current passes through path (A), we develop a voltage across the inductance of path (A). If this voltage is small, we haven't modified the amount of current flowing from the power system. If this voltage is large, we have so much VCC noise at the chip that it may not function. Catch-22. You can't win.

                              +----------+
                           ###|          |###
                              |          |
                           ###|          |###
        vias     (A)--->      |          |
Vcc plane X=====###========###| VCC      |###
                | |           |          |
Gnd plane X=====###========###| GND      |###
                              |          |
               bypass      ###|          |###
              capacitor       +----------+
                                  I.C.

If you wish to prevent a noise chip from interfering with the reset of your system, here is a better approach:

The Vcc plane (A) connects through inductor (B) to a bypass capacitor (C). The bypass capacitor has a short-wide connection (aspect ratio 1:1 or less) to the VCC pin (D).

The Gnd plane connects through a via directly to the bypass capacitor (E).

The Gnd plane connects through a via directly to the chip GND pin (F).

                              +----------+
                           ###|          |###
                              |          |
                           ###|          |###
          (A)    (B)   (C)    | (D)      |
Vcc plane  X==== L ====###=###| VCC      |###
                       | |    |          |
          Gnd plane  X=### ###|=X GND    |###
                       (E)    | (F)      |
                           ###|          |###
                              +----------+
                                  I.C.

This arrangement reduces noise flow in both directions, from the system into the chip, and from the chip into the system.

Ground bounce is conquered because I have provided a low impedance via from the GND pin (F) to the Ground plane.

Vcc bounce is conquered because I have provided a low impedance path from the VCC pin (D), through the bypass capacitor, to its ground via (E).

Best regards,
Dr. Howard Johnson

Errors-To: si-admin@silab.Eng.Sun.COM
Date: Mon, 28 Apr 1997 12:31:14 -0700
From: Larry.Smith@Eng.Sun.COM (Larry Smith)
To: si-list@silab.Eng.Sun.COM, gdpeter@sandia.gov
Subject: Re: Power/ground connections/bypassing on ICs
X-Sun-Charset: US-ASCII

Gary:

Some of us out here really do methodically design decoupling caps into the product. The methodology was developed and published by:

Larry D Smith,
Decoupling Capacitor Calculations for CMOS Circuits,
Electrical Performance of Electrical Packages Conference,
Monterey CA, Nov 1994,
Pages 101-105.

Since the methodology was published, it has been used it on several products. Lab measurements with a spectrum analyzer indicate greatly reduced power supply noise at many frequencies. Also, EMC/EMI radiation has been measured as much as 10 dB down after using the methodology. BTW, that guy from IBM works for Sun now... :)

regards,
Larry D Smith
Sun Microsystems