Attainable PWM accuracy?

Put the digipot on the bottom? GH

Cute circuit. However, this is clearly second order filter so fair comparizon would be with cascade of two RC steps. If I did my calculations correctly for R1 = R2 and C1 = C2 and assuming infinite impedance load Woodward's circuit has the same response as two step RC cascade. RC cascade does not need extra inverter, so it is not clear for me if there are reasons to prefer Woodward's circuit.

In some sense optimal tradeoff between filtering ripple and response time is given by Gaussian filter. Multistep RC cascade give good approximation to Gaussian, so in general it is not clear if more fancy (than RC cascade) filtering has advantages for PWM. Phil plans to use three stage RC cascade which already should be pretty good.

On Oct 11, 2019, Phil Hobbs wrote (in article):

about delta-sigma conversion, it occurs to me that one can improve rejection of power supply ripple by using an explicit charge-pump circuit in a delta-sigma DAC.

Joe Gwinn

Nice $5 power supply for a single white LED in there too, lol

Use switched filtering?

We'd have to bitbang that, for one thing, and for another there would be different numbers of transitions per cycle, which would hurt the linearity on account of the slew artifacts.

Cheers

Phil Hobbs

That sort of thing is sometimes done with ordinary APDs, but in this instance we're down in the guts of a SEM chamber with lots of other things going on. We have to worry about running into other detectors, and of course all SEMs of my acquaintance have video cameras inside, which require light sources....

Digital pots are generally pretty crappy--unless you go for the expensive ones that use metal resistors, the end-to-end resistance tolerance is typically 30%, with a poor tempco. Thus you have to use them ratiometrically and can't easily put a resistor in series to reduce the adjustment range.

Floating them at some weird voltage is perfectly possible but requires a fair number of additional parts. (One of our products has a set-and-forget laser current limit using a nonvolatile dpot that gets programmed at test time using a laptop--the dpot hangs off a +14V rail, but the laptop is floating, so neither it nor the dpot notices. If we were in mass production, it would need something like a Bus Pirate and a USB isolator.)

Cheers

Phil Hobbs

One approach to the PWM ripple problem is to use a S/H to sample the filter output at the beginning of the PWM cycle. The subtle problem is that even if the higher harmonics of the PWM frequency have died away adequately, the phase of the fundamental ripple component depends on the PWM code, which hurts the linearity.

Cheers

Phil Hobbs

Um then have a PWM gated constant current source feeding a capacitor, followed by a s/h.

Yeah, there are lots of ways to do it if you don't mind a bunch of extra parts.

What I wound up with is a vanilla 12-bit PWM driving a SN74LVC1G04 inverter that's powered by a REF5030 3-ppm/K reference, followed by a three-stage RC filter.

It's sort of funny, putting in a $5 reference just to drive a five-cent logic inverter, but that's where error budgets take you sometimes. ;)

Cheers

Phil Hobbs

Makes a lot of sense.

It might be worth exploring ways to put a zero in the filter response, at the PWM frequency. Maybe need less poles then.

Not sure what you were answering to, but the main and vernier PWMs could both be setup for a constant number of transitions.

isn't something like this

What I've never understood about delta-sigma is that you can make 25% duty cycle from

but one has twice as many transitions as the other. Rise and fall edges aren't perfect, so they won't make the same average voltage. Since the codes vary in density with the output voltage, that would seem to cause a linearity error and some excess noise. PWM should be better because it always has two transitions per cycle.

I guess the people who design ICs can apply tricks to fix that to 20 or 24 bit level. I suspect that, using discrete parts, we can't.

Looks like the best sub-10 PPM ref that we use is ADR363B, 3v 9 PPM, $1.89.

iptechnology.com:

n:

oing a

es from

but it

atively

It'll

om a

n output

ring to

r so, so

hought

a PWM, anyway?

will ease

are, but

xtension

think

at.

can

hey

s are

bits

a delta-sigma modulator usually have (should have) a much more "chaotic" output so it averages out

constant input to a delta-sigma modulator is "dangerous" because it can get stuck in some periodic pattern that result in tones in the pass band

and it is really hard to analyse/predict because it is not a linear system

But a mid duty cycle has more adjacent bits without transitions than a low or high duty cycle, hence the linearity error. I think.

I posted an LT Spice 2nd order delta-sigma modulator here a while back, but I didn't analyze it much beyond some FFTs.

n:

oing a

es from

but it

atively

It'll

om a

n output

ring to

r so, so

hought

a PWM, anyway?

will ease

are, but

xtension

think

at.

can

hey

s are

bits

Direct PWM may give better linearity perhaps, but worse noise because of th e frequency distribution.

TI Filterpro includes an active low-pass filter prototype that's a Bessel/Butterworth hybrid (I call it the "Bustleworth") that I like a lot for filtering PWM, which provides a good compromise between the linear phase response of the Bessel and the steeper slope of the Butterworth.