Sharing vias on decoupling caps

I saw a nice report from Analog Devices where they analyzed some filter topologies. As I recall, in the tiny cap + via structure, the via is almost as important as the cap.

If you have a BGA package and your caps aren't tiny and lurking on the opposite side of the board in amongst the power and ground pins, I'm not sure what good 400 of them are going to do you...

This consultant guy is clearly in the grips of some progressive brain disease. I offered to review the design for you, and our stuff is known to work.

John

Perhaps he charges $100 per cap he recommends, then gets a $1 kickback from the cap manufacturer?

Which, at really high speeds, is the huge low-Z, low-Q transmission-line structure formed by the power and ground planes themselves. The caps just help with the slow stuff.

Free, in fact.

Roger 1,2.

John

Gotta agree (and yes, I am back at last)

For the OP - a via through a 2.4 mm (0.1") board is about 1nH or thereabouts, as is 0.2 inch of 10 thou track. The parasitics of the device are almost as bad.

As to your :: So when you look at the overall picture, aren't vias shared anyways since current will flow from the 400 cap vias, through the 100 package pins, through the chip, back out the 100 package pins, back to the 400 cap vias? No reply yet from the guy.

::

Not quite. Current will return through the path of least impedance (real important at high speed) which *at* high speed means the path of least inductance usually. So the via that is closest to the power/ground pin will have the most return current in it.

Ya know, you can get real advice here for a LOT less than that conslutant (not a typo) who does indeed seem in the grip of some brain disease, or perhaps terminal currency attraction) if you can filter the noise :)

Cheers

PeteS

Some clarification

Return currents will take the lowest energy state return, which (as above) is usually the path of lowest inductance at high speed. If you need help / review I would suggest John is fully qualified and lower cost than your consultant. Of course, I would be willing too (hey, it's work).

Bottom line is

1. Your consultant seems to be full of shit
2. You need a real high speed expert to review your stuff

Cheers

PeteS

Given the money involved, why not build a test board and lease a network analyzer. I believe I mentioned this in your old thread where I had a case that I needed to know the inductance of the bond wires and built a test chip to measure it. Each pass of even a moderately simple chip is a few hundred $k, so the money was justified.

On your test board, try a few bypass schemes. Share the vias, don't share, spread out the cap values, blah blah blah. You can also try a few layouts with the chip in question. That is, not all the support circuitry, but just the high speed beast.

As an aside, generally the chip that takes the least user support wins. This includes external components. What invariably happens with products that are touchy is the initial product with the problem child chip gets out the door, then some time down the line, somebody in purchasing changes something and before you know it, the thing doesn't work. This even happens in chips. I had a case where the fab changed epi vendors and the THD of the chip skyrocketed.

more vias = less total inductance, as they are like they are in parallell, of course it would be better to have no vias at all - ive tried putting caps through holes in the pcb so each end cap is soldered directly to the plane itself.

I gues the right combination of cap values and trace inductance can give a wide spread of series resonances, this would advocate one via per cap.

Colin =^.^=

I have some interesting pictures of return currents in planes that I'll post sometime. For the OP, planes aren't solid layers at this speed - they are simply signal layers that happen to have copper between the signals.

Good for you

So many call themselves experts - so few are.

Cheers

PeteS

What the hell is a "shared via"? Give each cap its own vias?

robert