Reason for Ground Splits

Is your PCB power plane sliced into a bunch of little competing regions, all chopped up, like a map of the Balkans? I know people that have ten, fifteen, or more voltages on the same card. The proliferation of different power supply voltages has gotten completely out of control.

To help me better understand how power systems function, I have undertaken some measurements on working FPGA designs, and would like to share with you the results of that work.

The original web-based seminars on this subject have been taken offline, but the notes from my original presentation are here: Virtex-4 Power System Performance The notes cover the structure and purpose of a pcb power system, power system measurement and simulation techniques, and analysis of the simulation results. I prepared these notes with help from Mark Alexander at Xilinx, and Michael Brenneman at Ansoft.  

 

Reason for Ground Splits

Elya  Joffe writes with a question about grounding:

I am in the midst of writing a book on Grounding. The book is entitled "The Grounds for Grounding" and it will be published by IEEE Press and Wiley.

 One of the central chapters in this book is that related to grounding practices on PCBs. Of course, we know that there is no "real" ground on PCBs, but the issue of discussion is that of return and reference planes, chassis planes, etc.

One of the key issues I intend to discuss there is the issue of ADC/DAC grounding. That is not a simple issue… [I thought] the consensus was that when mixed A and D circuits are used on the same PCB, the best approach for grounding practices on the PCB is to use one SOLID and COMMON ground plane on the PCB, and only make sure that the routing of the traces does not lead to any conflict or crosstalk between traces.

In your article "Multiple ADC Grounding ", I was surprised to see that actually you recommend to use common planes when LOW RESOLUTION ADCs are used, but to split the planes (and to stitch to chassis) when high-resolution devices are used.

 From many discussions I have held, I found out that in most cases, a common plane is used and recommended.

 How can this be settled? If splitting the planes is recommended, are there any common recommendations or "common practices" as to the grounding approach, like tying the D ground pin of the device to the DGND plane and A ground pin to the AGND, and whether to bridge these grounds, not bridge them, etc....

I would appreciate your insight into this, particularly settling the dilemma I have regarding your previous article.

All the very best,
Elya

If you are not familiar with them already, check out these three background articles before proceeding:

  • ADC Grounding   (predecessor to the article you mention)
    adcgrounding.htm  
  • Multiple ADC Grounding   (which you already found)
    multipleadc.htm  
  • Common-mode Ground Currents  (presents a terrific visualization of the problem)
    ../news/7_02.htm

Regarding your specific question let me suggest to you an application that would require separation of the ground plane. Suppose you receive an audio signal from Eric Clapton's guitar. The signal amplitude is 100-mV p-p. Your job is to build a 24-bit studio-quality A/D converter. One bit of quantization noise in this converter equals 5.9 nV (referenced to the front end). On the same card you have a large processor that draws 10 amps.

When the processor starts and stops, DC current from the processor surges through the ground system. The DC resistance required to limit a current of 10 amps to a stray voltage less than 5.9 nV would be 0.59 nano-ohms. The common DC resistance shared between the processor and analog area must be less than this value. If you figure out how to make this work on one card with a shared ground plane (perfectly square, no cuts or jumps), please write to me and tell me how.

 Even with only 16 bits (considered inadequate for high-fidelity audio) you would need a common impedance coupling of less than 0.152 micro-ohms, a level I claim may still be unattainable.

 At 8 bits, you can get by with 39 micro-ohms. The end-to-end resistance of a 10x10 inch hunk of copper, 0.0013 in. thick, is .0005 ohms, at least in the right ballpark, although 10x too high. You might arrange the layout so that the input signal goes straight to the audio amplifier and stays away from the processor, and arrange the power connections so that all the DC current feeding the CPU doesn't plow straight through the analog area. You'll probably ground the input stage with a screw to the chassis right near the point where the signal comes in. You can get this to work.

Henry Ott's book, "Noise Reduction Techniques in Electronic Systems", goes into a lot of detail about these sorts of common-impedance coupling problems. It is a fast read, and it will kick-start your understanding of noise-coupling effects. If you do not have his book, get it.

I fully appreciate the pressures of book publication, and hope that you are able to complete your project successfully.

Best Regards,
Dr. Howard Johnson