Render and GPU Choices

Flame Hardware Questions
1

Doing some thinking about what GPU to choose for Flame.

One consideration is single GPU vs. multiple GPUs. And when thinking about multiple GPUs, trading off ML performance vs. background renders.

To baseline that, currently, how many folks are using just the standard Flame foreground renders vs. distributed rendering?

Currently using:

  • Standard Flame (foreground only)
  • Flame + Background Reactor
  • Flame + Burn Nodes
0 voters

I you pick your next GPU, assuming your system can support it (enough power, enough lanes, enough space):

Optimize for:

  • Any good GPU, doing normal renders only
  • Dual GPU configurations for background renders
  • Single large GPU - trading off ML performance over background renders
0 voters

These tradeoffs come into focus as we optimize GPU selection not just for renders, but also how well they handle the various ML features. In particular the Blackwell card is available in a 300W version, useful if you want to run two in the same system, and a 600W version which skews more towards ML tools loads.

2

I’ve never had Background Reactor be useful. If I try to keep working while having it render even the simplest timeline stuff it generally silently crashes, forcing me to fg render what it should have been doing.

How long are your renders? Except on the truly insane shots that only come around once in a while, foreground rendering has been plenty fast for me. I had a dual GPU machine for a while, and the 2nd GPU mostly just sat there unless I needed to make revisions to a bunch of heavy comps in a hurry - say when supervised by an agency art director, in which case it was worth its weight in gold.

Now that my workstation is at home, I did take a day and dual-boot my 4090-based gaming pc into Rocky and get Burn going on it. It was a significant amount of effort to get going (boy howdy does NFS admin suck on the Mac), and is now another install I have to maintain. I haven’t bothered putting any OFX plugins on it yet because I just can’t be arsed to deal with the licensing stuff. It was awfully nice when I had to crank out 80 shots worth of copy+paste paint fixes, though.

If it was me in your shoes I would want the fastest Nvidia GPU I could afford for Flame etc and then do all the ML stuff on another dedicated machine.

3

We use Burn extensively. There is a major caveat with it, that was the genesis of You’ve Been Burn’d

4

That’s my inclination as well. But always good to sanity check assumptions or workflow blindspots.

5

Interesting situation.

While inconvenient as a user, it’s not illogical from a software architecture perspective to treat each render as separate operation.

I’m sure you’ve sorted out a work-around since. One logical path would be to schedule the 2nd render as a separate operation, and have it read the main render as input, instead of using the processing stack as the input. A tad bit more scheduling logistics, but avoids double processing.

6

Like so many in-effecient operations in Flame, we just throw more metal at the problem.

When you render locally in Flame, the pipeline is optimized though, it concatenates the operations. Only in Burn does it re-render everything for every Render Node.