2080ti vs 1080ti : almost no performance upgrade for TFD simulation?

Well, everything’s in the title…. I have 2 of both cards in my setup, ran the exact same simulation twice on each card, so 8 runs in total :
1080ti : 44min and 52secs (average, one of the cards is primary display)
2080ti : 43min and 18secs (average, computing cards)

It is also worth mentioning that GPU load during the simulation is quite low, 15-20% on average on the 1080ti, rarely peaking around 45%, 11-14% on the 2080, didn’t see it peak past 20%.
Is this underexploitation normal ? I mean, in the TFD optimization guideline I read it is advised to have 2 gpus, a weaker one as primary display and a stronger one for computing…… this does make sense in a gpu renderer…. for instance, I use redshift and in some scenes the rendertimes are actually smaller when I do NOT use my primary display card for rendering. However, with TFD only using 10-20% of the card…… I tried using my display 1080ti and my computing 2080ti and got almost no difference.

I plan on swapping my 2nd (computing) 1080ti for a 3090 (when they are available again), this sure will help with redshift but…. from the looks of it i suspect this won’t help TFD in any way other than increased VRAM :frowning:

For the records, I’m using :
last TFD version v1-0_1465
C4D R23
up-to-date Nvidia Studio drivers 461.40

I shall leave this one for The Guru as I really do not have have any experience with the 2000 series. My 1070 is just about as fast as my Titan for simming but with the extra 4 GB RAM on the Titan I always use that. I do however regularly go well over 12 GB in sims so fall back to the CPU and as I am quite a fan of Arnold I think my priority will be to get more CPU power. Arnold can render TFD on the GPU if converted to vdb but I find the workflow so straightforward I tend to stick with CPU rendering. It just sits so well with TFD for shading. bcf files are a big slowdown for RS and you really need to convert to vbd to get them to render at a reasonable rate, Arnold plays with bcf very well.

As an aside…it cannot be easy coding this to work on the GPU’s for simming, no other sim software can do it…not even FumeFX

Could you provide a log file for simulation of a few frames on each GPU? You can attach it here or send it to support@jawset.com.
Please make sure you run the sims first, then send the file %APPDATA%\jawset\turbulence.log. This log file has detailed timings that will help to identify the bottleneck.

A possible caus is that either cache storage or complex C4D-side animation is slow enough that the GPUs are waiting for these stages to complete. A good test scene for a quick benchmark is the oil_well.c4d example scene from the plugin package. You can also run this scene in interactive mode to disable the cache.

What are you using to monitor GPU usage ?. Task manager is useless…use GPU-Z

Task Manager works if you select the CUDA graph:

True forgot that so used to having GPU-Z up !!

But it does not show utilization accurately

Thank you Jascha for your input. I’m in the middle of a production rush right now but I’ll dive into these tests / benchmarks asap and come back to you.
Like Paul, I tend to use GPUz to monitor gpu usage, seems to me to be more accurate than task manager.