Powering up analog part of the video card

Question

I am designing the video card for the device mainly powered by the +5 V of the ATX power supply. The digital part of the video card is also powered by the +5 V supply, as well as analog part. Analog part consists of several opamps, comparator, and some decision making logic (frequency selection PIC, LM1881).

My experience that 5 V power rail may be relatively contaminated to cause problems with the image, and I want to minimize these issues at the design stage. I have previously asked a number of related questions, and this one is about general architecture of powering the analog part.

What are the options for powering the analog part of the video card, and what are pros and cons?

Here are my thoughts so far:

1. Use regular +5 V power coming through the edge connector of the video card, and filter it using capacitors and ferrite bead. These +5 V may come long way to the card with a number of consumers on it, thus must be relatively noisy. Filtering may help a little, but LC filtering had failed the simulation;

2. Use other, more cleaner source of power. As an example +12 V, which is not that used in the system. And now another dilemma, how to step down from 12V to 5V:

   • Use simple LDO. Current is expected to be within 500 mA (maximum, to be safe, nominal must be 250 mA), and it seems to be a killer for the LDO device (3.5 W will require heat sink);
   • Use cascaded LDO, 12->9 then 9->5. This way there will be two devices dissipating the heat, approximately 2W each. I was told/read that such architecture may cause instability and noise if designed improperly (what is proper design here?);
   • Use cascaded DC-DC (12->6) and precision LDO (6->5). In this architecture device will have additional noise from the DC-DC converter, and I am not sure if it is better than just using +5 V from the system.

3. Use separate 5 V powering cord from the power suppy. ATX PSU is having a number of wires to power devices within the chassis, and I theoretically can make additional power connector for the analog part so that PSU would feed the circuits directly. There're clearly some kind of such architectures exist, e.g. modern high-performance GPU cards, but I do not know how they are designed. And I have two issues with this:

   • video processor and some circuits are not intended for that separate analog and digital power, thus I am not sure what will happen in case analog part of the circuit will accidentally be left without power connection - and if device may be damaged by the resulting currents;
   • by connecting another PSU wire I will create huge ground loop. I am not able to separate analog and digital grounds in the video processor by its design, thus it will be the same ground wire coming through edge connector and additional power connector.

Any suggestions basing on your knowledge and experience? Any success or failure stories to keep in mind for the decision?

Answer 1

I'd go with:

Use cascaded DC-DC (12->6) and precision LDO (6->5). In this architecture device will have additional noise from the DC-DC converter, and I am not sure if it is better than just using +5 V from the system.

Sure, a switch-mode power supply adds a new source of noise – but that's a source that you can control! Typically, what you'd try is to put that noise spectrally where it doesn't hurt that much (e.g. in an audio application, that'd be above 100 kHz), but in a video application, that might be hard – SMPSes work somewhere between 20 kHz and 4 MHz, and your signal of interest sadly usually spans both.
So, you'd often mitigate the noise by using a slow switching speed, or even spread-spectrum supply, which doesn't reduce the noise power, but smears (spreads) it over a larger spectral range. You'd then passively filter as well as feasible (i.e. here's where solid LC filtering comes in handy), and:

Follow that with a good linear regulator. Don't be satisfied with a LM7805 or alike. Use an LDO whose datasheet actually states noise rejection over frequencies, or shows the output spectrum for a given white input noise. Good LDO output buffering always helps – don't skip on the output cap (typically, use a tantalum + ceramic with each >= 1 µF each), and on the decoupling caps on every single supply pin on your analog part.