Switched-cap filters: Fallen from grace?

Used one about 7 years ago. According to their marketeers and confirmed by my scope and sweeper it made a nice filter. As chance would have it, I bought my first spectrum analyser the following week. Never touched these filters since :) john

With the advent of inexpensive programmable filters (based on DSP techniques) I think such things as the MF10 have only a niche market any more (if indeed they had any more).

Cheers

PeteS

I first learned about switched-cap filters in graduate school from Gabor Temes himself, and I was surprised that they'd never been mentioned in my undergraduate studies some years back. As far as I can tell (and others have mentioned), these days there are relatively few applications where DSPs don't do as good of a job, and even in cases where a switched-cap filter might have been cheaper, the lack of knowledge about them tends to still make people choose DSPs.

I do think that immediately jumping to a DSP -- that all come with hardware multipliers -- tends to make people lose sight of some clever techniques such as designing filters using only 0, +1, and -1 as the taps. That approach struck me as having good applicability to using low-end microcontrollers to implement filters, although when I looked at it a little more closely it appeared that wave filters might do just as well.

---Joel

Hello Pete, Hello John,

I can't say that I was ever dissatisfied except for rising chip prices. In medical ultrasound Doppler units they really had to perform and they did. Of course you have to provide good filtering to remove the clock or stuff might alias badly in the following stages.

DSP are still really expensive. Also, they are power hungry and not well suited for small battery operated gear. When you need a bank of steep filters you can quickly run out of horse power unless you'd use a big fat Blackfin. But that would slurp up the battery in no time. A uC might fit in as well but the minute you want a HW multiplier in there the price can skyrocket.

While an external DAC may not be needed an ADC usually is. For some reason most DSP won't go above 12 bits if they even have one on board.

The best thing I ever used a LMF100 for was a SSB demodulator for an optical instrument. I got > 60 dB other-sideband rejection with no tweaks. It's a very nice part in many ways.

On the other hand, it's noisy, noisy, noisy, even worse than op amp active filters, which is going some. That's the main reason I haven't used more of them.

Cheers,

Phil Hobbs

Hello Joel,

That's a point. Switch-cap filters haven't really been marketed well. When I did the last Doppler board with them the engineers at that client didn't know much about these filters.

DSPs aren't too great with battery gear plus most either do not contain an ADC or it's only 10-12 bits. Not enough for any serious audio stuff.

Well, if there only was a decent design software for wave digital filters. That could really boost uC sales but the mfgs don't realize it. And most people don't know about WDF either so it's unlikely there will be anything in the near future :-(

I suspect that's part of it -- the data sheets downplay the aliasing problem if it's mentioned at all, so folks will see some funny results. I remember having to work a bit to demonstrate how a switched-cap filter could alias a signal. Once we looked at the effort to pre filter the filter input, we decided to just sample the DSP a bit faster and be done.

That's a problem, no doubt.

The explanation I get from various companies' app engineers is that good processor processes aren't good ADC processes -- so if there's an ADC on there at all it's a huge compromise; even a 12-bit ADC may not have close to full performance -- which is why a grouchy old analog guy I sometimes work with describes the ADCs on the TMS320F2812 as "12 bit going on 8".

There is plenty of this stuff being used in ASICs, but presumably people don't think that it is the kind of standard product that you want to buy at the price at which they want to sell it to you.

Chris

Hello Tim,

Yes, you do have to pre-filter unless you can be sure there is no spectral part anywhere up there. One problem I sometimes saw when salvaging switch-cap designs that others had started was unfamiliarity with LC filtering. They tried to do it all with active filters and didn't realize that these can leak a lot. At one site they didn't even have inductors in the lab bins except for some huge powerline chokes.

Yeah but look at the MSP430F2013 and it's 16-bit ADC. Ok, it's slow and specsmanship on it is not great. IOW there ain't much in specs for use above 100Hz because they only had the powerline metering market in mind. But it performs remarkably well. Even comes with a PGA, courtesy of TI.

The Cypress PSOC devices are full of switched capacitor filters and other discrete-time circuit blocks.

Hello Matthew,

Thanks, will give these another look. It's been a while since I checked out PSoCs and back then they were quite expensive and the analog sections weren't really floating my boat. Could have changed.

Imagine analog design evolving to programming chips in the analog equivalent of VHDL.

John

I designed them at Maxim, Exar, and Reticon (yes, NMOS). They are good in ASICs and customs, where you sell/buy the system on a chip. Customers that can "see" the filter complain about the noise, distortion, and charge injection, but in the ASIC, they just see the net system.

If the circuit is on the back end of the filter is on chip, you can get around the charge injection problem if the back end circuit is itself a sampled data system. That is, you make it synchronous with the clock. Many of the off the shelf SCFs had built in divide by two circuits that made synchronization impossible since you don't know the state of the divider.

One thing also to consider in filtering is the objective. That is, you are trying to get rid of some signal. But if the distortion artifacts exceeds the attenuation of the offending signal, what have you accomplished? (Throw in PSRR in the mix). On ASICs, you would cut back the filtering to the point where the offending signal is a bit beneath the noise. Fine for a standard product, but tell the customer their design isn't optimal and many balked at the notion of changing it. On a standard product ASIC, this wasn't an issue. If you look at the early modem chips, the filtering was reduced to reach the optimal level rather than follow the Bell design.

Getting back to PSSR, few customers really appreciated what a crappy filter they built with those MF10s. The PSSR is very frequency dependent, and worse with high Q. To get a really good SCF, it had to be leap frog based, which is just impossible with a building block design, i.e. biquardratic sections.

Yup. They alias power supply noise right down into the signal.

John

In article , John Larkin wrote: [....]

(Shutter)

pinout of component D1 is begin pin anode : copper with tin; pin cathode : copper with tin; end pinout;

value of component D1 is begin Silicon diode type 1N914; end value;

.....

D1 map pins (anode -> TheAC, Cathode -> TheDc);

C1 map pins (PlusEnd -> TheDc, MinusEnd -> CircuitGround);

.........

That was much easier than:

TheAc -->!---+--- TheDc ! --- --- ! GND

:)

How about:

Run the DC-DC at the same frequency as the clock of the switched Cap, or some harmonic there of.

Why do we still use schematics to design PC boards? When I tell people that I use schematics to design logic, they think I'm a dinosaur.

John

It demands on the complexity. VHDL has some quite nice properties on large projects. Verbose, certainly. ...and by design. Ten years ago I was a schematic hold-out. I rather like VHDL now, at least for the guts.

Certain DC/DC designs lend themselves to be used with external clocks, but the issue in not limited to switching supplies. Power supplies are never perfect zero-ohm sources, so their will be noise on them just from the circuit operation. Say the circuit works on band splitting audio, as in the simpler modem designs. Then the transmit channel circuitry will be generating noise right where the receive filter is trying to elliminate it. If the PSSR of the receive filter in the region of the transmit spectrum is poor, you may have a situation where the PSSR related noise is worse than what gets past the filter.

I've seen SCF designs with gain from the power suppies if measured at center frequency of a high Q resonator.