rf - At what frequencies does PCB design get tricky?

I have designed many mixed-signal PCB's where the highest-frequency component is the microcontroller's crystal oscillator itself. I understand the standard best practices: short traces, ground planes, decoupling caps, guard rings, shielding traces, etc.

I've also put together a few RF circuits, at 2.4GHz and ~6.5GHz ultra-wide band. I have a working understanding of characteristic impedance, ground stitching, balanced vs unbalanced RF feed lines, and impedance matching. I've always contracted an RF engineer to analyze and fine-tune these designs.

What I don't understand is where one realm starts to cross over into the next. My current project has a 20MHz SPI bus shared between four devices, which has let me to this question. But, I'm really looking for general guidelines.

1. 
   

   Are there guidelines as far as trace length vs frequency? I assume that ~3 inch traces are fine with 20MHz (15 meters), but what is the general case?

   
   


2. 
   
   

   As frequencies increase, how to prevent long traces from radiating? Are striplines and coax the way to go?

   
   


3. 
   

   What is the RF characteristic impedance of a typical microcontroller output stage, anyway?

   
   


4. 
   

   etc.

   
   

Please feel free to tell me anything I'm missing :)

Answer

1. Are there guidelines as far as trace length vs frequency? I assume that ~3 inch traces are fine with 20MHz (15 meters), but what is the general case?

At my work, the guideline is, if the electrical length of a trace is longer then 1/10 wavelength, you need to treat it as a transmission line. At a minimum, this means you must terminate with a resistor matched to the impedance of the line. How do you figure out what resistor value to use? You estimate what the impedance will be during design, and then you adjust the value to minimize ringing during DVT.

Now, there is some subtlety here about the true meaning of 1/10 wavelength. For a sinewave, this is straightforward. For a square wave, which is the sum of many sines, you must use highest frequency component as your estimator. As you sharpen the corners of the square with a faster slew rate, you increase the frequency of the fastest sine competent.

What this means is, for a digital signal, drive strength directly affects the electrical length of the line. Higher drive strength can easily turn a line that does not ring into one that does.

I learned this the hard way when a supplier made an "improvement" to a digital buffer without telling us. This change increased the slew rate, which caused ring so bad that the receiving chip started to latchup. A board we produced that had been working fine for years suddenly started randomly locking up.