Electronic – Chosing XTAL Capacitors for Microcontroller

My favorite electronics book is "High Speed Digital Design: A Handbook Of Black Magic". I highly recommend this book. It seems expensive, but it is totally worth the money. This book has 12 pages on choosing a bypass cap! The author, Howard Johnson, also teaches some classes with decoupling caps as one of the topics.

Some important things that I've learned over the years, and have been backed up by this book, is that the "standard practices" with decoupling caps are almost always wrong and there is more art than science when it comes to choosing and routing them.

There are lots of calculations that you can do regarding decoupling caps, but much of those are not accurate due to many things. The caps themselves are vary wildly (especially the higher dielectric caps like X7R). The PCB layout changes things greatly (and you'll need to think in 3-D for this one). Temperature and voltage will change the behavior of the caps. A single cap will behave as both a "power supply smoothing cap" and a "AC signal return bypass cap". Etc.

What Johnson did was, after a lot of experimentation, figure out that inductance is the most important factor and it swamps almost every other consideration. So the goal when selecting and placing decoupling caps is to use a lot of physically small caps, with the highest practical value, and route them so the total inductance is as low as possible.

The ideal would be to use lots of 0.1 uF caps in an 0402 package. Place them under the chip on the back side of the PCB. The cap be routed as in the image below. And the vias go directly to the power/ground planes (not to the chip's power pins, as that would usually increase the inductance). If you place the cap under the chip then sometimes you could share the same via without any issues.

The reason why a 0.1 uF cap was chosen is because it is the highest practical in an 0402 package. The reason why 0402 was chosen is because it is the smallest practical size, and you want to use a lot of them to get the effective ESL/ESR down. Of course all bets are off if you have a 2 layer PCB without power and ground planes.

I don't want to belittle the use of the math, that is important, but the complexity of power supply decoupling and AC return paths often makes the math not so practical in the real world. In the real world, a "rule of thumb" really helps. Of the many rules of thumb for this topic, it has only been Howard Johnson that has proven the other rules don't work and provided this better rule. My experimentation and experiences has shown this to be true.

My seat-of-the-pants understanding for load capacitors (corrections invited) goes like this:

When a crystal is cut for a certain load capacitance, it is measured with that capacitance across it during final factory trimming. There is nothing magical about the value. It is simply a way of saying, that if you design your circuit to present that same capacitance, then your crystal will be within the stated (.005% or whatever) tolerance.

So, you add up all the capacitance in your circuit, and then add in what's needed to bring it up to the spec. We'll use your numbers. The stray capacitance due to the traces on the board obviously will vary with the board, so let's guess 1.3 pf. A number I made up, to go with the capacitance of the microprocessor's oscillator, stated to be 1.7 pf. So, we've got 3 pf in parallel with the crystal. The crystal wants 18pf, so we have to make up the 15 pf difference with discrete parts.

Since the two load capacitors are in series (Gnd->cap->xtal->cap->Gnd), we double the cap value to 30pf. Two 30 pf caps in series give us the 15 pf we're looking for.

Note 1. I tried searching for typical PCB stray capacitance. It was all over the map. Suffice it to say, that as the hardware gets smaller, the capacitance will keep getting smaller. A lot of typical values claimed less than 1 pf.

Note 2. If there is more capacitance than spec, the crystal will oscillate at a lower frequency than specified. If there's less, then it's higher. You can see, that if you want to trim the oscillator to spec, it's easier to shoot for a lower capacitance and add some later, than to try the opposite.

Note 3. For fun, look up "gimmick capacitor".

Note 4. My "seat of the pants" explanation is sufficient as an introduction, and this technique works in many cases, but not everywhere. For a more in-depth look at the EE principles behind those capacitors, see this answer.