pick-and-place

From the prespective of powering chips, it pretty much is.

A tapered transmission line is not inherently lossy. But, as the line impedance drops, things like skin resistance do increase losses. A 1 ohm transmission line made of copper and FR4 is pretty lossy.

Right. Bypass caps help for slow (nanosecond) transients and gross load steps.

Interestingly, many/most big CMOS chips, like processors and FPGAs, have surprising amounts of on-chip bypassing, and many add on-package caps as well. An Altera Arria II GX95 has half a uF of on-chip capacitance on the 0.9 volt core.

Thanks for the update... care to elucidate?

Your speculation is at odds with the state of the practice as it regards PCB power plane modeling and performance prediction.

We'll never get a coherent explanation from Larkin because he's just a mechanic following published guidelines, my suspicion is that the 330nF caps serve as distributive lossy plane loading to kill resonances.

ote:

nslation for you, sheesh. Isn't the SRF of the 330nF 0603 rather low for by pass?

mpedance does. And the impedance in the frequency range that mainly matters is dominated by ESL, which is about the same for all 0603 ceramic caps.

u to know as it translates directly into supported rise times and damping.

MHz or 500 kHz. The high-frequency impedance will be about the same.

lmost any ole ceramic caps here and there.

the caps in the case of power plane layouts. Looking into the parallel com bination of capacitor and transmission line power plane, the capacitor dive rts the high frequency current loading off the line thereby decoupling the bypassed device from other devices on the same power plane feed. This is on ly good within the working frequency range of the cap, so a 100KHz or a 1GH z SRF almost certainly does make a difference in that regard. Unless you're working at a few tens of GHz, there's no wavefront hitting capacitor, it i s a lumped element. That 330n with max 10MHz SRF is only slightly better th an worthless. A 3.3n would do much better.

is interesting. I remember looking at the power planes as a transmission l ine. The effective impedance decreases with radial distance so the amplitu de of the wave decreases as it spreads out and diminishes the need for deco upling caps. But if there are a lot of chips on the board all clocking at the same time I expect the power planes won't be enough for most designs.

pair as a parallel-plate capacitor, with a small amount of inductance from the vias that you need to get at the planes from the surface.

ig, rectangular, fairly lossy transmission line. If you can access it throu gh an embedded SMA test connector, it TDRs that way: atransmission line of an ohm or two, with weak, fuzzy reflections off the edges of the board. As you add bypass caps here and there, the structure starts looking almost lik e a perfect capacitor. Bypass cap SRF is not visible in the TDR measurement of plane impedance.

to do with the ineffective low SRF (faux) decoupling caps preventing the h igh frequency switching currents from launching wave energy out onto the pl ane, and this at frequencies most likely to couple through the Vcc/Gnd poin ts of the other ICs connected into the plane.

on't design electronics.

- he just tinkers with stuff until it works. His enthusiasm for ignoring a capacitor's series resonant frequency (SRF) in favour of its equivalent ser ies inductance (ESL) which he believes to be the same for all 0603 packaged capacitors is indicative of his level of design insight. If you don't thin k hard about what you are doing you do tend to see things as a bit simpler than they really are. Sadly, you can get away with it most of time - but I' ve spent quite a lot of time cleaning up after people who didn't.

I'm pretty sure Larkin has it wrong. The capacitors don't decouple the acti ve device from the power plane in the usual sense, they decouple the power planes from the active device by providing a conjugate match to the plane i mpedance thereby damping/killing the amplitude of reflections from incoming waves on the plane.

I'm not speculating, I'm building electronics and selling it. What do you do for a living?

Follow published guidelines? Not me! I'm on record here, for years, of mocking the Howard Johnson-type, multiple capacitor value, SRF Spice model orthodoxy. I use a tenth, sometimes a 50th, of the caps that are recommended in some FPGA appnotes. And my stuff works.

A close power:ground plane structure, seeded with a few 330n 0603s, is an excellent power system. If actually measuring plane impedances, and measuring power plane noise on real boards, make me a mechanic, then there should be more mechanics and fewer theorists in this business.

It sounds like you are agreeing that my products work, and are now struggling to explain why. Enjoy.

You can encase the columns in steel shells. Rather expensive though. If the shells are slightly oversize you can pump grout to fill them up. Something you can look into if you want.

?-)

Seriously, I never struggle...

What blather.

The bypass caps furnish power to the chips in the mid-frequency range, where the power regulators are too slow and the copper planes can't store enough energy. They also lower the high frequency plane impedance a bit, if there are enough of them close to big fast loads.

I don't know why so many people get worked up over bypassing. It's easy and almost always works first try. I haven't had a bypassing problem on a digital board in a decade or three.

Wideband analog designs can have problems if you're not careful. Some opamps have PSRRs near unity, sometimes worse, so a little supply noise can cause big trouble in a high-gain, wideband amplifier circuit. That's about the only place where ferrite beads or resistors belong in power distribution systems.

We already poured three big concrete footings, each with a steel frame that goes all the way to the roof. The flat roof and floors are now structural plywood, bolted to the steel and to the side walls. Everything of significant mass (shelves, equipment) is bolted down. It may be totalled in the next big quake, but it probably won't collapse and crush us all to death.

People want to knock it down and build a 40-story condo anyhow. Residential stuff here is going for $1000 per square foot or more.

ote:

ote:

translation for you, sheesh. Isn't the SRF of the 330nF 0603 rather low for bypass?

; impedance does. And the impedance in the frequency range that mainly matt ers is dominated by ESL, which is about the same for all 0603 ceramic caps.

now.

you to know as it translates directly into supported rise times and dampin g.

ond

h almost any ole ceramic caps here and there.

of the caps in the case of power plane layouts. Looking into the parallel combination of capacitor and transmission line power plane, the capacitor d iverts the high frequency current loading off the line thereby decoupling t he bypassed device from other devices on the same power plane feed. This is only good within the working frequency range of the cap, so a 100KHz or a

nt is interesting. I remember looking at the power planes as a transmissio n line. The effective impedance decreases with radial distance so the ampl itude of the wave decreases as it spreads out and diminishes the need for d ecoupling caps. But if there are a lot of chips on the board all clocking at the same time I expect the power planes won't be enough for most designs .

ne pair as a parallel-plate capacitor, with a small amount of inductance fr om the vias that you need to get at the planes from the surface.

a big, rectangular, fairly lossy transmission line. If you can access it th rough an embedded SMA test connector, it TDRs that way: atransmission line of an ohm or two, with weak, fuzzy reflections off the edges of the board. As you add bypass caps here and there, the structure starts looking almost like a perfect capacitor. Bypass cap SRF is not visible in the TDR measurem ent of plane impedance.

ing to do with the ineffective low SRF (faux) decoupling caps preventing th e high frequency switching currents from launching wave energy out onto the plane, and this at frequencies most likely to couple through the Vcc/Gnd p oints of the other ICs connected into the plane.

u don't design electronics.

cs - he just tinkers with stuff until it works. His enthusiasm for ignoring a capacitor's series resonant frequency (SRF) in favour of its equivalent series inductance (ESL) which he believes to be the same for all 0603 packa ged capacitors is indicative of his level of design insight. If you don't t hink hard about what you are doing you do tend to see things as a bit simpl er than they really are. Sadly, you can get away with it most of time - but I've spent quite a lot of time cleaning up after people who didn't.

ctive device from the power plane in the usual sense, they decouple the pow er planes from the active device by providing a conjugate match to the plan e impedance thereby damping/killing the amplitude of reflections from incom ing waves on the plane.

No kidding, Sherlock, they can't help but do that. But the RF conjugate mat ch is what gets you maybe 40dB attenuation of power plane disturbance noise at the capacitor. You're really quite dull sometimes.

mocking the Howard Johnson-type, multiple capacitor value, SRF Spice model orthodoxy. I use a tenth, sometimes a 50th, of the caps that are recommende d in some FPGA appnotes. And my stuff works.

Trust John Larkin to be sufficiently out of touch to think that Howard W. J ohnson represents orthodoxy.

Howard W. Johnson and his half-baked "High-Speed Digital Design - a Handboo k of Black Magic" was the first heavily publicised text on high-speed digit al design to hit the market (in 1993), but it was ill-thought-out rubbish. Most of the relevant phenomena get mentioned, but the text lacks coherent o rganisation and is remarkably unhelpful.

I got more - earlier - from the Motorola, Fairchild and GigagBit Logic appl ication notes, backed up after 1990 by Peter C.L. Yip's "High-Frequency Cir cuit Design and Measurements" ISBN )-412-34160-3 and "Microwave Components, Devices and Circuits" ISBN 0-471-91277-8, neither of which is all that goo d, if a whole lot better than Howard W. Johnson's contribution and a whole lot less pretentious.

Sadly, Howard W. Johnson is the guy who went around giving expensive semina rs all over place, so even John Larkin got to hear about him.

an excellent power system. If actually measuring plane impedances, and measuring power plane noise on real boards, make me a mechanic, then there should be more mechanics and fewer theorists in this business.

If you understood what you measured rather better, this might even be true.

struggling to explain why. Enjoy.

I don't think Fred has any trouble understanding why your products work - i f you fiddle with something for long enough, most people can get a good eno ugh intuitive understanding of what is going on for most practical purposes .

A more coherent theoretical understanding of what you were doing and why it worked would probably have given you better products faster, but good enou gh can make money. In a niche market, it can make quite a bit, for a while.

ote:

ote:

ctive device from the power plane in the usual sense, they decouple the pow er planes from the active device by providing a conjugate match to the plan e impedance thereby damping/killing the amplitude of reflections from incom ing waves on the plane.

I don't understand it either, but I doubt that it's blather.

where the power regulators are too slow and the copper planes can't store e nough energy. They also lower the high frequency plane impedance a bit, if there are enough of them close to big fast loads.

Transmission line impedance is L/C^0.5, so the extra capacitance can lower it quite a bit, as long as the wavelengths involved are appreciably longer than the inter-capacitor spacing. At 1GHz in free air a quarter wavelength is about three inches - dropping to two inches in regular dielectrics. ESL complicates the picture. but not a lot.

nd almost always works first try. I haven't had a bypassing problem on a di gital board in a decade or three.

Because you are under-ambitious?

I've run into minor problems about once per decade. Once we'd realised that it was a by-passing problem it was always easy to fix.

opamps have PSRRs near unity, sometimes worse, so a little supply noise can cause big trouble in a high-gain, wideband amplifier circuit. That's about the only place where ferrite beads or resistors belong in power distribution systems.

I had a circuit where feedback from the rails could have been a problem - a nd it did oscillate, but not because I'd got the - elaborate - by-passing w rong. TI had been less than candid about the input impedance of their CMOS- input op-amp, which just did DC-level setting. It was easy to stop the osci llation, but the compensating capacitor became the single through-hole comp onent on a surface mount board that otherwise worked straight off the drawi ng board. I was livid.

things never happen at the same time. clocks take time to slew and to propogate.

I'm not speculating, I'm calculating. Didn't you see those formulas? ;)

You were claiming that SRF in and of itself made a difference to the high frequency behaviour of a bypass cap, whereas it's really just the ESL of the cap plus vias, as I demonstrated with a simple calculation. The notion that _reducing_ the value of a SMT bypass cap without changing its size or shape could somehow improve bypassing is pure cargo cult engineering.

You didn't look at his picture very carefully. That's a bare board. And why would 330 nF ceramic caps form "a distributed lossy plane", whatever that is?

And that paper has serious problems. Their experiment vs theory "agreement" is off by an order of magnitude, for one thing, and for another, there's no reason to use some cheesy transmission line approximation when there are free codes available that calculate it as accurately as you like.

You might notice that the JL's measurement of the impedance at early times (i.e. high frequency) is approximately flat at 1 ohm, which is well within the calculated range. Their calculated curves are actually much more believable than their data. The big series resonance dip _improves_ the bypassing, and all those little resonance peaks don't do a whole lot--it averages out to an ohm or less, which is right in the range of the TDR measurement. The coupling between ground pins of different chips will be a great deal smaller than the self-impedance.

Cheers

Phil Hobbs

wideband

(There is that narrow-band improvement close to the SRF.)

On 09/01/2014 11:43 PM, John Larkin wrote: > The bypass caps furnish power to the chips in the mid-frequency > range, where the power regulators are too slow and the copper planes > can't store enough energy. They also lower the high frequency plane > impedance a bit, if there are enough of them close to big fast > loads. > > I don't know why so many people get worked up over bypassing. It's > easy and almost always works first try. I haven't had a bypassing > problem on a digital board in a decade or three. > > Wideband analog designs can have problems if you're not careful. > Some opamps have PSRRs near unity, sometimes worse, so a little > supply noise can cause big trouble in a high-gain, wideband > amplifier circuit. That's about the only place where ferrite beads or > resistors belong in power distribution systems.

You said it. You really start to notice the via inductance when you try bypassing a multi-gigahertz wideband amp. Mere paranoia doesn't begin to cover the level of care needed.

My first try at the pHEMT/SiGe:C NPN cascode trick oscillated at 12 GHz. At the time, I didn't have a spectrum analyzer that would go that high, so I measured the frequency with a manual wavemeter: I measured the DC offset while moving my hand up and down over the offending part of the circuit. The amplitude went smoothly up and down depending on the separation, repeating itself over a spatial period of about half an inch.

A wideband ferrite bead (Murate BLM18B series) and several 0402 bypasses in parallel fixed it.

Cheers

Phil Hobbs

And your point is? Remember that engineering is all about getting something close enough.

Topside power pours and lots of vias!

We almost always add 0402 gate resistors to PHEMTs, but it's tragic to toss all that bandwidth and noise performance to buy mere stability.

And most serious chips are clocked from internal PLLs anyhow.