Arnold GPU Render Times Increasing Every Frame
I'm seeing pretty consistent render times in my test scenes, although they are not particularly complex or heavy. Behavior could depend on many things, including the Arnold version, graphics card, driver version, etc. Maybe a specific object or setting in the scene is triggering it. To get more detailed information on the issue, could you please render a sequence with Arnold debug log enabled and send us the log files? You can setup logging in the render settings on the Diagnostics tab.
It looks like you might be running out of GPU memory. When this happens Windows will move some of the data off the GPU. This is nice because you don't crash. But it makes things way slower. The log files show how much GPU memory is available at the start of the render and then how much we were able to get during the render:
log 0106
00:00:00 22030MB | GPU 0: NVIDIA GeForce RTX 4070 @ 2490MHz (compute 8.9) with 12281MB (7074MB available) (NVLink:0)
00:04:02 32138MB | peak GPU memory consumed 6958.21MB
log 0108
00:00:00 25612MB | GPU 0: NVIDIA GeForce RTX 4070 @ 2490MHz (compute 8.9) with 12281MB (10126MB available) (NVLink:0)
00:01:24 31866MB | peak GPU memory consumed 10062.74MB
Notice how the 0106 render had 7GB available instead of 10GB from the 0108 render, and it uses pretty much all of that 7GB while the 0108 used all of the 10GB it had available. I suspect you probably need 12GB or so of available memory to properly fit everything on the GPU. I went ahead and graphed the info from all your logs and it does show a strong correlation with render time and amount of free memory at start of render:
The next question is why does your GPU sometimes have 7GB free and sometimes 10GB? This could very well be a bug in Arnold, though that would be a bit surprising because I would expect a buggy Arnold to either always hold on to the memory, so available memory stays constant or progressively gets worse over each render. Instead it seems to fluctuate and then at a certain point (frame 108) it somehow resets itself and you're back to 10GB free. Another possibility is that this could be c4d consuming GPU memory for its own use. A third possibility is that you have some unrelated program also running that is consuming varying amounts of memory?
We maybe never noticed a slowdown because we have GPUs with more memory and never ran out of memory. But we should be able to detect this slow reduction in available memory. We'll check that out in case that lets us repro.
If you have multiple GPUs, dedicating a GPU for the renderer would be ideal since then it would have the full 12GB of memory to work with while the c4d and other applications can use the other GPU...
