delta-sigma again

Even older FPGAs. lol

Multiple stages of decimation seems like multiple opportunities for aliasing!

Multipliers and MACs are free nowadays. My RC-RC IIR model can be done with two MACs.

I designed a series of electric power meters (for utility end-use studies) that sampled the AC line at 27 Hz! 16 CT channels on an MC6802 slow 8-bit processor.

I even acquired waveforms of a sort by equivalent-time sampling.

Just because

mandag den 18. november 2019 kl. 16.25.23 UTC+1 skrev snipped-for-privacy@highlandsniptechnology.com:

that's what the filters are for

just say it out loud; the industry standard, know to work, used in millions of ADCs way cant' be used, because you didn't come up with it ;)

Every decimation resamples the signal at a different frequency.

I've seen clocked filters (like switched cap filters) do terrible aliasing things.

What I don't understand is why people do complex things when simple things look like they will work. Specifically, why DSP guys don't seem to approve of simple integrator-based IIR filters.

mandag den 18. november 2019 kl. 21.00.01 UTC+1 skrev John Larkin:

after a filter that band limits the signal so there is no aliasing

apples and bananas

they did the math and realised that a simple integrator-based IIR filter is sorta ok for smoothing data, but for decimation CIC and FIR is much better and computationally efficient

IIR filters are not capable of being tuned to null a spur. The FIR coefficients, if chosen correctly, DO have a lot of flexibility in their filter characteristics. For one, the transient response of a six-element FIR goes to zero at time-sample #7 after the transient.

Mathematically, the FIR is clean. If you don't do math, that isn't obvious.

Absolutly true. You have grasped the fundamental principle of multirate filtering.

Ok, let us know how it works out for you.

ptechnology.com:

nd

so

AMs

Anyone seeing aliasing in a switched capacitor filter doesn't know how to u se a switched capacitor filter. Literally.

ons of ADCs way cant' be used, because you didn't come up with it ;)

That has already been explained to you. IIR filters are not so simple if y ou design them rigorously, meaning making certain they are stable rather th an just designing something and seeing if it *seems* to work. But we know that is how you often work when dealing with things you don't fully underst and.

I think John invented a form of exponential smoothing and is amazed with its performance. Which is truly glorious, small number of these sections can do miracles if the coefficients are properly tweaked. And they can be made inherently stable. But their pulse response is a bit weird. They are very useful, though, e.g. as the lowpass filter of a digital PLL.

Best regards, Piotr

An RC filter on the input to the ADC isn't all *that* expensive.

People generally do more complex things because the simple things that look like they will work go horribly wrong under some circumstances.

It is well known in numerical analysis circles that modelling an exponential decay differential equation is a harder problem than it looks because at some stage the exponentially growing parasitic component of the solution will rise up and bite you in the backside. Rounding error can cause a lot of trouble with feedback.

That is one reason why people are wary of IIR the way you intend to use it.

I meant the amount of remaining resources would be enough for anything else. I remember my experiments with Cyclone some 20 years ago, an IQ mixer and decimating filters occupied most of the device. The multiplier alone took 1/3 of the resources. Could probably have been optimized usig some clever tricks, but today one just doesn't care. I have recently put some serious multichannel megasample-level processing pipeline into Cyclone V and used maybe 15% of its resources. Pity there are no gull-wing variants anymore, I don't trust the BGAs.

Best regards, Piotr

In 2008 I designed a board for a major networking company that does DSP on a time code signal using an FPGA from around year 2005. That design was vi able then and the QFP100 fit the tiny board size just fine. Still making t he board although the last remaining inventory of the product is at Arrow a nd they are starting to raise the price.

I've been exploring some alternatives and found Lattice has an XO3 in the Q FP100 along with three Chinese startup companies. I'm trying to research t he potential longevity these companies might have. Right now all three see m a bit immature. Gowin seems to be the most responsive so far. I'm suppo sed to have a phone call with them tomorrow.

BGA solder joints seem to be a lot more reliable than gull wings. And one advantage is in inspection: you can't really inspect the BGAs!

Well, just a little.

I'm going the other way. I have a 20 mbps delta-sigma data stream that I want to convert to, say, 16 bit parallel form with good signal properties, as one feedback in a control loop. I don't want aliases, delays, or weird phase behavior.

An RC lowpsss model is simple and stable, if done as a simple integrator with a bit of negative feedback. But I see people making an integrator out of a pile of sections and a sheet full of math. Curious.

OUT = OUT + K * (IN - OUT)

is the RC model, where K is small. Tau is just the clock period/K. Sometimes the multiply by K can be just a right shift, but MACs are free nowadays so there's no reason to do that any more.

I don't see any exponentially growing parasitic component. Every time I did it, it just works.

It only works nicely for a high ratio of clock to filter cutoff, 10:1 or better. The step response is obviously steppy if you take big steps. The impulse response is infinite exponentially descending steps to zero.

Filtering a delta-sigma stream is a big ratio of clock to cutoff, several hundred in my current case, 20 mbps stream and maybe 50 KHz signal bandwidth.

BGAs are quite reliable. Wetable flank QFNs are quite nice packages too. BGAs and anything with a center pad need X-ray inspection, though.

Do people use xrays to inspect production boards? I've been told that it's rarely done.

We do optical inspection, but that's more for process control than per-unit QC.

Not so. If you're concerned about reliability you do. In the aerospace and automotive markets they do.

It's probably good enough for your environment. The issue with x-ray inspections of central pads is paste coverage. If your process isn't perfect there will often be voids in the paste, increasing the thermal resistance accordingly. The proper collapse of the solder balls is the obvious problem with BGAs. The edge ones you show are perfect but there really isn't any way of knowing what happened in the center without x-ray inspection.