This Spring I've been teaching my new Advanced High-Speed Signal Propagation course at sites all over the country. In addition, I still maintain and teach the material in my original course, High-Speed Digital Design. Thanks to all who have written in with questions and discussion about these courses (and many other things).
One of the many topics I like to discuss in the original course is grounding, especially the topic of cutting up your ground plane. I avoid cutting the ground unless it is absolutely necessary. This article presents an interesting way to visualize the benefits (and pitfalls) associated with the practice of cutting up your ground plane.
Common Mode Ground Currents
You've probably measured noise voltages at various points on a digital ground plane (ground your probe to the chassis and touch the tip to your digital ground plane). This easily demonstrable concept seems paradoxically at odds with the general concept of a ground region as a continuous, solid metal object.
Most digital engineers think of their digital ground plane as a solid sheet. Unfortunately, this mental analogy makes it difficult to visualize certain classes of common-mode noise problems, the sort of problems that crop up when trying to isolate sensitive analog circuitry from big, bad digital processors.
To get around this conceptual difficulty I'm going to suggest an alternate way to visualize ground regions. Instead of thinking of your digital ground region as a solid sheet, I'd like you to draw a picture frame (Figure 1). Now connect a current source from side to side across the picture frame. The current source represents the aggregation of all the signal currents flowing through traces on the board, all of which ultimately flow in through the ground (or power) planes back to their respective sources.
In your drawing, the current source drives current into the right side of the frame. The current splits, returning across the top and bottom of the frame, rejoining at the left side to re-enter the current source.
Shown the basic picture-frame model, most engineers immediately recognize that large currents surging through the non-zero impedance of the frame will cause noise voltages measurable at various points around frame (NOTE 1). For some reason, it seems easier to see this with a frame than with a solid plane, which is the whole value of this type of analogy. To make the picture-frame analogy work, you must postulate that the impedance of the frame satisfies two constraints:
- It is sufficiently low to limit ground noise to a level acceptable for your digital logic (tens of millivolts), but
- It is at the same time sufficiently large to permit the accumulation of voltages damaging to highly sensitive analog circuitry (microvolts).
Since there is a factor of 10,000 to 1 difference between these levels, it is not too difficult to imagine a frame that might satisfy both constraints (NOTE 2).
The picture-frame model neatly explains the concept behind single-point grounding. As long as the picture frame connects to one and only one other object, in one and only one place, no current can possibly enter or leave the frame (NOTE 3). To see this, add to your drawing a second frame below the first, and connect them with line A (Figure 2).
Current only flows where there exist both outbound and return pathways. Current will not go from the first frame, through line A to the second, and then return to the original frame along the same path (that's ridiculous). The second region remains isolated from the first (and vice-versa). Current is not exchanged between the two regions.
An isolated ground region must have only one connection to the outside world. Not two, and not more than two. Not three, either. Just one. As soon as you connect to the frame in two locations, some of the current circulating in the frame will escape into the outside world.
To see why, add a second connection (line B) between the two frames. As long as the connections on the first frame are made at two distinct points, there will be voltage differences between those two points. These differences, however small, force current out line A, returning on line B (Figure 3).
At the same time, currents generated within the second frame can circulate through the first. If you are trying to isolate the frames from each other, the second connection defeats your purpose.
Let's apply the picture-frame concept to a real product architecture. Suppose your product comprises an analog region and a digital region, and suppose further that your tolerance for noise in the analog region is 10,000 times less than the tolerance for noise in the digital region (i.e., you need 80 dB isolation between the two regions).
As a first step towards your conceptual understanding of this problem, draw an analog frame and a digital frame, side by side, connected together at one and only one point (analog on the left, digital on the right, connection in the middle). Since there is only one connection between the two frames, currents in one frame will not interact with currents in the other (Figure 4).
The gap between the two frames is called a "moat", and the single connection the "drawbridge". The drawbridge is a perfect place to locate your A/D or D/A converter .
Of course, your digital region probably needs to connect to other stuff, and the interface is likely a single-ended ground-referenced connection, so the digital ground frame requires a connection to the chassis. Draw the chassis surrounding both regions, and pencil in a connection on the right side of your diagram between the digital frame and the chassis (Figure 5).
Since the analog region still connects with one and only one drawbridge to the rest of the world, it remains isolated from everything else (NOTE 3).
What goes horribly wrong, in the classic evolution of a dysfunctional product architecture, is that someone decides to incorporate a chassis-referenced analog input on the left side of the picture. This new input requires a connection between the chassis and the analog frame (shown in green, Figure 6).
Unfortunately, any ground connection you pencil in from the chassis to the left side of the analog region represents a terrible noise-management mistake. Immediately as the connection is made, currents from the digital frame can flow out the right side of the digital region, through the chassis, and back in the left side of the analog frame, passing through the chassis-to-analog-ground connection (shown in green). These currents ravage your sensitive analog circuitry as they pass through the analog frame on their way back to the digital frame.
Remember that most of the digital return current still flows through the digital frame, as shown in Figure 4 (but not shown here). The only current shown in Figure 5 is the stray current that leaks out of the digital frame, passing through the analog frame. The proportion of the total digital current that flows through the analog frame is small, but even one part in 10,000 (according to the initial design assumptions of this article) can cause difficulty.
As an aside, I should mention that many products work perfectly well with ground-referenced connections on both sides of the analog region exactly as drawn in Figure 5. These products function for the simple reason that, even though stray digital currents circulate freely in the analog region, the stray currents are sufficiently small that they do not bother the analog circuits. Many analog large-signal devices (meaning hundreds of millivolts) work fine in the presence of modest amounts of stray digital current. For example, a PLL clock recovery circuit, a transmit oscillator, and a serializer chip all work with signals large enough that you needn't worry about stray digital currents. For these types of chips, simply moving the sensitive circuit away from the digital region (about 1/2 inch) and providing locally filtered power provides enough isolation for a large-signal analog device to work perfectly. If you see such a circuit implemented on an isolated analog region, with a moat and TWO drawbridges, you should know that it works not because of the tricky ground-plane configuration, but because the circuit wasn't very sensitive to noise in the first place. Most circuits that work with TWO drawbridges would work just as well with a single contiguous ground region. Try it . It is when working with truly sensitive analog circuits, like radio receivers or sensitive photo-detectors, that the extraordinary measure of moating pays significant benefits.
Going back to the architecture in Figure 6, to reduce the magnitude of bothersome circulating currents you have three alternatives:
- Change the topology of the circuit,
- Put a high impedance in series with the unwanted current to block its path, or
- Put a low-impedance in shunt with some part of the circuit, encouraging current to flow in that direction.
These three generic solutions suggest three possible ways to modify your system to get the isolation you may need.
I have listed the best alternative first. Changing the topology of the system avoids the whole problem. The basic idea is that current leaves or enters a frame when you connect to it at two different spots. That suggests that it may be OK to connect to two other things as long as you do so from the same point on the frame. This idea is explored further in .
Idea number 2 suggests placing an opto-isolator, transformer, GMR isolator, or at least a common-mode choke in series with the circulating path. These types of isolation devices convey information without passing ground-to-ground currents from side to side across the interface. Referring to figure 5, an isolation device can be placed in series with the analog input, between the analog and digital sections, or at the digital-to-chassis connection, whichever is easiest. A cheesy alternative to these good solutions is to use a digital differential connection across the analog-to-digital drawbridge. The differential link does not provide as much isolation as the other measures, but is sometimes good enough to function.
Idea number 3 involves bolting a solid metal sheet in shunt with the digital ground plane, hoping to divert current through the impedance of the solid sheet instead of through the more circuitous pathway involving the analog region. This approach only works to the extent that you can attach a solid sheet with an impedance lower than the already low impedance from side to side across the digital ground region itself. You will not get many dB of improvement from this method, but if you only need a few dB, it can help. Adding a second solid metal sheet in shunt with the analog ground region provides an additional measure of improvement .
So far, we have discussed the first-order effects of moats and drawbridges. We must now discuss two more subtle effects—resonance and parasitic capacitance.
Any time you weakly connect ground regions there exists the possibility of creating a gigantic, uncontrolled resonance. Investigate resonances by stimulating each plane region with a network analyzer and measuring the response at various other points around the system. Whether a resonance causes trouble depends on whether the resonant frequency lies within the band used by your circuits. Modifications involving the length and width of drawbridge connections, and the number of type of screws used to connect to the chassis, have a major effect on common-mode plane resonances.
Most resonances can be easily defeated by eliminating the moats, coupling the whole system together with one large continuous ground region. If you are working with large-signal analog devices, the one-solid-ground approach is the best strategy, as it avoids entirely the possibility of common-mode resonance between sections, yet provides enough isolation for most large-signal devices. Only implement the moat-and-drawbridge structure when you are convinced it is necessary .
The parasitic capacitance between surface-layer traces on your board and your chassis drives significant amounts of current into the chassis whether you have established an intentional metallic connection there or not. This "invisible connection" acts as an unavoidable second drawbridge between your digital region and the chassis. During debug, simply moving your board further from (or closer to) the chassis will change the nature of this coupling. If you discover that moving the board affects your noise performance, the only solution is to keep the board away from the chassis, or cover up the surface-layer traces with grounded metal.
Dr. Howard Johnson
In your system, the reference-plane regions will look like solid sheets (full of holes), not frames. They may be connected at certain points (drawbridges) and separated at other spots (moats). Whenever you see such an architecture the first thing I want you to do is identify how many drawbridges connect to each section, and make a frame-based drawing to represent the interconnections. Next, look to see if it incorporates any of the noise-reducing ideas 1-3.
NOTE 1: Most people correctly guess it is the inductance of the frame members, not their resistance, that generates most of the noise. If you widen one of the frame members until it approaches the width of your ground plane, and move it directly underneath the noise source, you have created a microstrip geometry. The inductance mechanism applies to this situation just as it does to the frame—the voltage noise you observe on a practical digital ground plane results from the returning signal currents working through the finite inductance of the ground plane itself.
NOTE 3: This statement assumes there are no parasitic connections between the ground region and the rest of the circuit. Parasitic connections, such as the parasitic capacitance of the traces and chips on your board to the chassis, can be quite significant in low-level noise problems.