First, I had a short look at the LM329az. Over EUR10 at TME.eu, and obsolete according to digikey
I have done some simulations of a cost effective PWM and it works as intended.
1. Multiple phase shifted PWM outputs of a micrcocontroller (3V3)
2. 74HC541 which is powered by a voltage reference.
3. Summing / averaging resistors. (A few k Ohm, exact resistance not important).
4. Analog filter (Fet opamp, no DC leakage).
This setup has huge advantages.
1. When combining 3 outputs (for example from a motor timer in a uC) then the effective output frequency is tripled because of the phase shifting.
2. Either resolution or output frequency can be increased because (Pulse width of each timer does not have to be the same)
3. The summing / averaging resistors already generate a much smoother output, even before the filter.
These three factors combine into an improvement of more then quadratic. You can choose whether to put this improvement in higher pwm frequency or more resolution. Same tradeoffs as always.
In the picture blow I also added a bigger resistor (R13, 22Meg) The Idea is that you can tweak the output with yet another PWM channel. It's tweaking range should be a few lsb of the other outputs combined, but it needs calibration in software.
And the zipped KiCad for the simulation. Simulation looks very promising. With 4 outputs I had a settling time of around 15ms (100MHz pwm) to ppm dc level. I ran out of resolution of the simulator. Resistor values have for the filter have not been optimized yet and need further tweaking. It has become quite clear that the dominating inaccuracies will be because of the switching times, imbalance of High / low of the TTL driver, voltage reference etc.
(Edit: Attachments and images are a real mess on this forum)