CPU v CUDA v RTX...trying to speed up this shot

[fixeighted]

A comparison...same slow animated camera in static scene, tested with cpu/cuda/rtx, using settings from cpu - 008noise, no time limit and default gpu settings when switching.
2080ti, 3950 amd, latest everything and recommended driver.

CPU - 8ish minutes
CUDA - 7.38 first frame, subsequent frames 11
RTX - gave up after 10 minutes as it was clearly going to take at least a half hour or more...just the precompile took an age.

Silly me for buying a 1000 quid gfx card
Good for games at least

[Muhammed_Hamed]

Hey fixeighted Thanks for the report

For optimal performance you need to use progressive mode with V-Ray GPU
You will need to use the least amount of render elements, it can affect performance a lot on GPU(we are working on this now)
If there is empty space in the frame or if there is bokeh noise, these are cases where V-Ray GPU takes longer than expected to render.

It is odd that RTX is slower than Cuda, we have been sending such cases to Nvidia. Is there a chance I can take a look at your scene?
FYI the 2080ti is 3x slower than a 3090ti, you will find 3090tis at less than 1000 Euros at the moment in Europe

When you say CPU, do you mean the standard V-Ray renderer? or Cuda x86 mode?

Best,
Muhammed

[fixeighted]

Thanks for the suggestions/info

Yeah I used progressive in all cases, with no dof, only denoiser.
There are some large maps but I made them smaller all so it would fit on the card.

When you say empty space, what do you mean?
This scene is a large terrain with an object that sometimes is the close subject and other times the terrain is mostly in shot.
There is sky and aerial perspective/clouds.
Neither viewpoint is quicker using gpu.
RTX is just annoyingly slow and every time I tried it since getting the card in 2019 it has been slower iirc.
Mostly, by the time it's chugged along compiling, compiling, doing whatever else it does before it starts, the cpu would have
been happily dealing with the lot and just rendering.
Ultimately it's no big deal and I'll revert once more to cpu simply to get the work done.

The comparison was standard cpu setup, which always works, against the other two.
In every scene I've tried it is always cpu that beats the card and honestly, who has time to mess around on a deadline?
I'd send the scene but I'm right in the middle of trying to fix the other weird stuff that's occurring, whilst trying to
get shots done, so after I fix everything I can pack it if you're interested...maybe it's user error....

[seandunderdale]

hey Muhammed, any reason to go for the vanilla 3090 over the 3090ti? Only reasons I can think of is price and maybe power usage?

[Muhammed_Hamed]

hey Muhammed, any reason to go for the vanilla 3090 over the 3090ti? Only reasons I can think of is price and maybe power usage?

Hey seandunderdale

The 3090 is within 0-3% of the 3090ti performance, specially the good models like the ROG Strix or MSI SuprimX
The 3090ti could be a lot more expensive, consumes a lot more wattage and produces a lot of heat. It is not worth as you don't get any extra performance in return

Best,
Muhammed

[Muhammed_Hamed]

When you say empty space, what do you mean?

Hello fixeighted

For example a car setting in an HDRI/Backplate setup, where big part of the frame is empty(backplate pixels that doesn't require any calculations), which reduces the GPU usage
On the other hand if you have an interior scene, the entire frame or most of that will push the GPU usage a lot higher. You can check that in AfterBurner or GPUZ, one scene will push your GPU to 280 Watts, and another scene will push it to only 150 Watts

There is sky and aerial perspective/clouds.

The sky is probably part of not pushing the GPU usage high enough, we are aware of that and there is effort now to make it better

RTX is just annoyingly slow and every time I tried it since getting the card in 2019 it has been slower iirc.

We have been sharing such cases with Nvidia, while RTX might not offer a speedup over Cuda it shouldn't be slower. If it is slower than Cuda, it is a corner case that needs to be investigated.
That being said, RTX shines when you have large surface of geometry that interacts with lighting, like leaves, grass, hair..etc
If the shading is heavy in the scene, with SSS, procedural effects and so on, RTX will not offer much of a speedup over Cuda but again it shouldn't be slower than Cuda

by the time it's chugged along compiling, compiling, doing whatever else it does before it starts, the cpu would have
been happily dealing with the lot and just rendering.

This is at the top of our current priority list at the moment, specially with scattering plugins. the compilation times are too long on GPU

In every scene I've tried it is always cpu that beats the card and honestly, who has time to mess around on a deadline?

Fair enough!
Here is a fun example from production if you have time, RTX is nearly 2x faster than Cuda and a lot faster than CPU
Most of the shading is baked into bitmaps, this gives RTX the edge as shading doesn't slow rendering much
Another point is how a 2080ti is 3x slower than a 3090, which will be 6x slower than a 4090. So your GPU is showing its age by now

I would still love to take a look at your scene if possible.
Thank you!

Best,
Muhammed