continuously variable Bessel filter

What could be any simpler then DSP or sw-cap ?

Digital pot is RC-RC-RC-RC... At 600kHz it is going to be a mess.

What are the THD, the accuracy and the dynamic range requirements?

That would give you 256 different frequencies.

What an insight.

Sometimes it seems to me that Thompson is your father and AlwaysWrong is your son... There are certain simularities...

VLV

John Larkin schrieb:

Hello,

have a look into the datasheets and aplications of the Dpots, here:

Bye

_I_ am insulted ;-) I'm sure that AlwaysWrong is insulted to :-)

But I have successfully designed digitally tunable active filters on a chip, up to 1MHz. Ask Mark/qrk.

But I certainly won't tell Larkin how to do it.

BTW, VLV, your ego doth approach that of Larkin. Careful there, boy! ...Jim Thompson

Do any swcap filters work at 600 KHz? But they are time quantized and noisy, and I need jitter in the 1 ns RMS range. And DSP is complicated and power-hungry, as mentioned.

Yeah, dpots are just a bit slow for this frequency range. They're intended for audio mostly, I guess.

Ultimately I'm going to find the peak, or maybe centroid, of a sort-of Gaussian pulse, to within a small number of nanoseconds. Gotta remove some noise first. Distortion maybe 0.1% or less. Pulse heights are pretty constant. Noise will probably be dominated by the physics itself.

Oh, yeah, I could scale the resistors in 2:1 steps and make a sort of resistive DAC. But that would take 8 analog switches, not an 8:1 mux, more cans but not too bad. Using a quad analog switch and five resistors gets me 16 steps, still not bad.

Do you use Spice to evaluate "simularities" ?

Hey, you've been helpful in spite of yourself!

John

Thanks for the link.

The parasitic capacitance is kind of high on that particular part. The only dpots that look like they might work in the 100's of KHz are the lower-valued ones, 1 or 2K. There seems to be a trend for the ones with low numbers of taps, like 64 or 100, to have less capacitance. That makes sense.

John

Does it have to be a causal filter?

We came up with a nice centroid-time-finder circuit that obviously can't be causal. The answer has to be delivered when the pulse is only half done. Reality sucks!

The fix for that one was to add an analog delay line to let various signals catch up with other signals, at the cost of some insertion delay, which we can accommodate elsewhere in the system. At roughly a couple MHz bandwidth and a couple of microseconds of delay, that delay line started to look like a project on its own.

But once you filter a signal medium-hard, a peak detector is almost the same as a centroid finder. All that convolution stuff. The time-machine-avoidance thing, and centroid balancing, has already been done in the filter.

Great fun, but the customer thinks we ought to do this for free, fuzzy spec feature creep, so we need to be pragmatic.

John

We have a box that's somewhat similar to that... it's an adjustable notch filter, from something like 100kHz to 1.5MHz. That's not nearly as fancy of a transfer function being implemented as your filter, of course, but it does work well for us. We use fixed C's and then the R's are effectively made from multiplying DACs (from Analog Devices), which we found to be better parasitic-wise than digipots.

Each unit does have to be calibrated insofar as DAC settings vs. desired notch frequency; this is done during production via a PC running some simple-minded-but-effective "autotuning" algorithm -- the results are stored in the internal flash ROM of a microcontroller. (This box was designed some years ago -- if we were doing it again today, I'd suggest having the board contain its own signal source and then performing an auto-tune every time the frequency was changed -- less work for production, and it automatically compensates for, e.g., dramatic temperature differences between our room temperature production environment and where the boxes might actually be used. We've also kicked around the ideal of just switching it to a DSP implementation as well, given that 16-bits at 1.5MHz is no longer that big a deal, but the volumes are low enough and the new design work large enough that it's not obvious that it'd be worth it.)

I sometimes think of small 8-bit micorcontrollers today as, "large flash memories with a bit of programmable logic included for free."

---Joel

Take a small FPGA. Make a basic Delta-Sigma ADC. Make a moving average filter with variable length. Make several stages of moving average if needed. Make a basic Sigma-Delta DAC. Here you go.

Vladimir Vassilevsky DSP and Mixed Signal Design Consultant

Is delta-sigma practical for a few-hundred-ns wide pulse?

Problem is, FPGA design resources are overloaded around here, and an FPGA needs program storage, a zillion noisy power supplies, ADC and DAC anti-alias filtering, JTAG, stuff like that. I need to filter the signal then find its peak to within a few ns, non-trivial to do digitally. I'd prefer some all-analog solution, easier to design, low power, and quiet.

We considered doing all this digitally (using a real ADC and DAC) but it's a lot of work.

John

AD has some "matrix switches" that you can individually turn on any combination of pins. Major limitation: supply voltage. See ADG738/739. Maybe a crosspoint switch? You haven't said what kind of frequency resolution you need.

Personally, I'd take a serious look at the multipliers, and definitely at the state variable implementation. Of course, if there are any MDacs with well characterized bandwidths (from the reference inputs) that would be good. Are there any?

Right. The dpots are good to around 5 MHz at best. ADI has an 8-bit MDAC, AD5450, for around $2, good to about 12 MHz. I could stick a bunch of them into some state-variable filters and program all of them from a single dip switch.

Yeah, a little ARM with flash is a pretty good deal for $4 or whetever. The ADC and DAC are free, too.

But I can make my customer look up the codes in a list and set the dip switch!

John

OK, but this is entirely different problem then.

As a contrary, doing it digitally is simple: interpolate the waveform.

So, it is down to the classic solution for finding the centroid of a pulse: lowpass filter, then differentiator, then comparator to zero ? There is not much benefit in making the lowpass adjustable; an octave gear shifting should be sufficient for practical purpose.

If you already have some basic design with DSP, then it would be no brainer adapting it to the problem.

Vladimir Vassilevsky DSP and Mixed Signal Design Consultant

We have some products where the contract says the RS-232 bit rate has to be DIP-switch settable, and shipped in the 19.2kbps position. We've shipped hundreds of these things, and I'm certain that no one has *ever* touched those switches after they left our building!

something like Spartan-3AN has build in flash for configuration, only needs two supplys 1.2 and 3.3, jtag is just 5 wires and a connector

think you could combine the filtering and center finding in one go

-Lasse

The particular customer that I'm working with assigns a zillion digit stock number to any FRU. A different number to every box that has different dipswitch settings.

Their document control system is called AGILE. They call us and ask us for files of theirs that we might have and they can't find. We call it FRAGILE.

John

Haha... that might encourage *me* to use as many DIP switches as possible. :-)

Apparently this one, now acquired by Oracle? -->

I bet it'll soon be even more fragile!

---Joel

One approach would be to use transconductance amplifiers cascaded with capacitive loads; the program current to the amplifiers modulates the output impedance and thus the bandpass. This definitely qualifies as a low power solution, you could do six poles with under 50 mA. See the datasheet for National Semi's LM13700, figures 11ff.

Yup. In one of our customer's annual reports is a statement something like "and we have almost recovered from installing the Oracle software."

Remember when HP almost went out of business installing Oracle?

John