Keyshot 9 GPU rendering on Quadro RTX5000 16Gb LOL

Started by sdesaulles, November 13, 2019, 04:11:48 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

sdesaulles

Well, I suppose I shouldn't have got my hopes up....
Having seen the awesome speed claims for KS on GPU rendering with RTX cards, I deliberately chose the most powerful GPU that I could get in a mobile workstation. A NVIDIA GTX 5000 with 16GB of punch....
I just downloaded KS9 today. Waiting for a license to be sent through but went with a watermarked trial in anticipation of some stunning "benchmark" camera file results.
KS8 on my HP Zbook was giving about 135FPS on the camera benchmark file, so I was anticipating big numbers on the GPU.

First attempt: GPU icon greyed out, thought I'd try a system re-boot....

Second try: Driver too old, so updated to the latest driver.

Third attempt : Driver (441.12) too new, there are "known issues" where KS pretends to run on GPU but is actually still on. KS9 wants 436.48 something that isn't available from Nvidia, I download the closest I can find, 436.30. It installs with afew warnings about overwriting a newer version.

Fourth attempt: Hoooray! KS9 finally has the GPU icon not greyed out, I select it and wait while KS9 progress bar slowly updates materials etc to work on GPU.

So, load the benchmark and watch that FPS fly! Now with a 16GB RTX card, running at a blistering 120 FPS, 15FPS slower than on the CPUs.

Doh! Worth the hype. Well, not for my configuration it would seem.

Who's seen the biggest jump then? (I'm interested in forward jumps mostly, but hoping someone has had a more disappointing result from a 16Gb RTX graphics card than I have to make me feel less humiliated about my choice of RTX for KeyShot9!)

Ho hum. KS X to look forward to with combined GPU and CPU? I won't hold my breath.

Eugen Fetsch

Can you compare the samples count and share the results (screen capture), please? Maybe there is a bug in the FPS counter.

soren

Please do not compare samples per second and fps. You cannot even meaningfully compare fps between interior mode and product mode on the cpu.

We could easily have made a GPU renderer running at a millions of samples per second but extremely high amounts of noise. The correct comparison is to let each renderer run until your reach a desired quality level and measure the time and compare that.

While comparing samples and fps is very convenient it is not a good metric. The samples per second can be used compare different gpu setups and fps can be used to compare different cpu setups - you cannot compare between them only based on the numbers.

Søren

mafrieger

#3
@Søren

thanks for this description!

What do you think of making some kind of "Automatic" for this:
QuoteThe correct comparison is to let each renderer run until your reach a desired quality level and measure the time and compare that.
e.g.
render till a special amount of noise is in the rendered image  (maybe plus another criteria)
of course this would be harder if denoise is activated...

In addition the result of some kind of automatic quality analyses could also be a stop criteria for normal render processes (non benchmark)...

edit: btw from technical pov this may be some how related to https://www.keyshot.com/forum/index.php?topic=25276.0 "noise layer/pass"

DriesV

I think if you look at the actual rendered image, you will find that GPU rendering clears up the noise much faster.
If you really want a hard metric, then you could perhaps compare number of samples at a given render time.

For the benchmark scene, I get these results:
CPU (32 threads): 128 samples in 10 seconds
GPU (RTX 2080, speed similar to Quadro RTX 5000 Mobile): 2000 samples in 10 seconds

I would encourage you to compare image quality though, and not get caught up in numbers when comparing CPU and GPU.

Dries

sdesaulles

Thanks for the replies. So, question, is the KeyShot Camera, camera_benchmark.bip still a benchmark or not when using KS9?

DriesV

Albeit simple, and arguably outdated, the camera scene can still be used as a benchmark. Faster CPUs and GPUs will get higher scores. Just don't compare FPS and Samples Per Sec when comparing CPU and GPU.

Dries

sdesaulles

#7
Ah, I'm clear, it is still a benchmark but I shouldbn't use it to compare CPU and GPU, so its, errm, not a benchmark for what I'm trying to test. Is there a better file to use?
Also, still waiting adfter 48hrs for my license code, so not a worry for the moment unless I want KeyShot watermarked accross all my renders. When I do get a code, hopefully this side of the weekend, I will run some timed renders on GPU and CPU and see what differences I get.

Thanks

Stephen

sleby

As I can understand you cannot directly compare cpu and gpu samples or fps. Why don't you simply set max time for GPU to 1 min and then try how long does it take for CPU to get to the same quality?

DerekCicero

This page shows the driver information:

https://www.keyshot.com/forum/index.php?topic=25224.msg105610#msg105610

which does reference Driver R435 U2 (436.30). I have updated it again to try and make it more helpful, but if there are ways to make it clearer, please let us know.

Unfortunately the issue of proper driver compatibility is going to come up as these RTX cards first come to market and get broader usage.