Fusion Performance Obeservations

So I recently purchased a AMD 7950x for my render node (32GB, RTX 3090, NVME SSDs) to go with my Workstation Threadripper 3970x (128GB, RTX 4090, NVME SSDs) and have made some very interesting observations.

When rendering some reasonably heavy 4k comps I am finding the 7950x is 30% faster than my Workstation. I am a little surprised by this as the 3970x over all has more processing bandwidth and memory and faster GPU (I have Fusion set to auto for GPU) but I have always noticed that Fusion seems to use one master thread that seems to max out and all the others bounce around at 0-80%. Because the single clock on the 7950x is so much higher I assume this main thread is able to push through more work, or it could be some optimisation that I am not seeing.

I have been able to really churn through a lot more work with this new setup, as I very rarely Render from my Workstation.

Very interesting report. I need help in thinking about a new machine.

Have you tried tuning the settings in the Memory tab of the Preferences? If you find that Fusion isn't using all of your processor cores efficiently, you can try turning on Simultaneous Branching and increasing the number of frames processed at once for final rendering.

Each individual branch of nodes must be processed serially--it can't be efficiently multi-threaded because the order of operations is important. However, with Simultaneous Branching turned on, Fusion is able to process each discrete branch on its own, which can improve utilization. It can also lead to instability on some systems and/or some comps, though, which is why it's off by default. Whether or not it makes a difference to your rendering speed depends on the structure of the individual comp. If it's just one long linear graph it won't do anything, but if you have lots of branching in the comp it could speed things up significantly.

Rendering more frames at once can also improve speed, but only to a point. You have to be careful not to run out of RAM because swapping to the drive, even an SSD, will probably slow the render down more than greater CPU utilization speeds it up. Sometimes cutting the rendering back to only 1 frame at a time can improve performance in a comp that's very high res or that uses a large number of buffers.

I used tuning the same and weighted by thread count as well, not much diff. I have done tuning previously and found that there is often not a big difference, in general most nodes are multithreaded anyways and the only times I have seen thing max out on all threads is when a single node is doing massive MT calculations. I do believe that a strong GPU has become very useful in Fusion, and have seen the GPU pushed hard sometimes, but its pretty difficult to gather any really useful metrics other than running a bunch of 'average' comp's and noting the duration.

In my experience Render X frames at once has little effect. Of course it depends on the structure of your comp. When Fusion was CPU-based, 4 was optimal for me. Recently I checked and it seems that 2 is optimal.

I tested again with the current composite. 2 frames were still the best. In reality, however, the difference is only about 1%. (Edit:Each time it is measured, there is a variation of about 1%.)

The GPU-based Fusion uses very little CPU. Does a higher clock CPU affect Fusion more than many core CPU?

That's good data. Thanks for doing the testing so I don't have to!

I do believe that a strong GPU has become very useful in Fusion, and have seen the GPU pushed hard sometimes, but its pretty difficult to gather any really useful metrics other than running a bunch of 'average' comp's and noting the duration.

For very small comps, it can make a huge difference. Like if you're just batch processing some footage to generate depth maps.

But for production shot comps, I typically have to disable GPU processing anyway to prevent crashing, so the benefit goes way down as the comp complexity increases.

I'd be curious as to what other people's experiences are with reasonably complex VFX comps.

Same experience as Chad.

The current beta seems a little more unstable. Especially the new tools like the magic mask struggle on my machine when I don’t purge the cache before using them.

For me interactivity is also better on bigger comps when disabling GPU. Sometimes when moving masks they simply don’t update correctly if I enable GPU processing.
This is regardless of resolution and also happens in FHD.
There is also this issue in certain setups where playing back a tool that’s supposedly cached to RAM doesn’t play back in real-time if it’s GPU processing is enabled.
But this mainly happens to me when using older
machines, when I work in a studio. I usually just disable the GPU processing altogether.

To be fair, my 8GBs of VRAM in my RTX 3070 are too little in many scenarios.

Maybe for smaller comps a faster single threaded cpu helps to push data faster to the GPU and back.

As far as this beta is concerned, I have discovered one thing.

In the above test, Final render consumed 9.4 GB of GPU memory in v18.5 beta3. v18.1.3 use 18.7GB.
Render time of v18.5 beta3 is 7m32s. v18.1.3 is 13m21s.

Beta almost twice as fast and half the memory?

Well, that is good news. I must admit I feel like I have been cranking through work and have had to close down Fusion less due to GPU memory getting exhausted (24GB). I have GPU set to auto and maybe have 2-3 crashes a day but that is giving Fusion a bit of a hammering and I used to have around 15-25 a day. Keep it up BMD!

I tested it on a small comp. v18.5 beta 3 was 1.5 times faster and GPU memory was still half. The more complex it becomes, the greater the difference may be.

I would like to use 18.5 once the OFX issue is resolved.

I wonder if because of the speed of these new GPU's there is now no need to Cache certain (GPU accelerated) tools?