CPU or GPU, I'm lost

[FranklinTre]

You're right Nicinus , just finished a project and it has been a pain in the a** to work with 8GB of VRAM. I could do pretty much everything with 24 GB... maybe 48 in the future.
Honestly, for the startup, what matters the most is the RAM and HDD speed, you still a huge amount of RAM depending on your scene. The part where the VRAM is exploited is not so much of a problem with a poor 1070 TI since Vray 5. The bandwitch of the 3090 is probably anihilating the problem even if you use 8K textures.

Forest and Itoo are not anymore problems IMAO. It used to be with previous versions. Though it uses a lot of VRAM, so in that regards it is still the main concern.

All of my problems when rendering in GPU nowadays comes with VRAM limitation, not anymore from the lack of software support.

[squintnic]

I think the ultra slow load times really hold it back - try IPR on a 3990x - its great
featuire parity must be addressed at some point imo

[oglu]

There is an feature comparison overview.
https://docs.chaosgroup.com/display/...orted+Features

For us the main two missing freatures are full featured Matte Shader and Clipping with Geo.

[roboxx]

Just setup a new workstation and was also curious about CPU vs GPU speed. Did some test render with a AMD 3970X vs RTX 3090 FE and the rendertimes were 12:50 min vs 10:50 min. on a quite simple scene in 5K wih some metal shader stuff. So it might be different on other scene scenarios, but the 3970X is a real beast and working in IPR mode is really straight forward and interactive.
I found it quite annoying in the past when working in GPU mode with 4 x 1080TI, because the whole loading and preparation of the scene before the GPU render started took quite a while, also had sometimes lots of refreshing and startovers when using GPU just through clicking in the viewport or opening some tabs or similar........
This might have changed with Vray5, but the high upgrade prices especially for the render nodes did hold me off for upgrading to V5.

[francomanko]

isnt the GPU loading issue being addressed?

[roboxx]

for me working on VrayNext 4 the issue is still there....and annoying

[squintnic]

isnt the GPU loading issue being addressed?

Is it? This would be great. Not seen any acknowledgment of this but might have missed.it.

roboxx I have a workstation with same CPU and GPU using vray 5 and the render times are a similar balance.

[francomanko]

im probably wrong tbh

[Micha]

GPU loading is like before, I don't see a difference.

[seandunderdale]

I spent 9 months working on GPU projects with dual 1080ti system having used CPU 100% before that. Im so glad to be back on CPU rendering. Even though my cpu is only a 1950x I feel I get loads more done. I spent so much time just trying to get the GPU to do the thing I wanted, and writing bug reports on here, or asking questions to colleagues about workarounds. Once I got the shot working, dual GPU (even 1080ti) + cpu in hybrid did render fast, but I lost that time in avoiding feature difference issues, or bugs. Back on CPU and I can crank shots out, even though the render times are a bit slower.

It feels so nice not having to worry about what effing driver I have installed for my GPU to get renders done.

Once I decide to upgrade my machine to a Zen 3 CPU I wont look back to GPU rendering at all. Not until Vray 6 at least.

[Alex_M]

seandunderdale theedge
theoratically, the RTX 3090 is supposedly 25 times faster than the 3960x.
That would also put the RTX 3090 10 times ahead of the 3990x, for a third of the price.

I'm pretty sure that RTX 3090 is not anywhere near 25 faster than the 32-core Threadripper 3960x or 10 times faster than the 64-core Threadripper 3990x, lol.

[FranklinTre]

I'm pretty sure that RTX 3090 is not anywhere near 25 faster than the 32-core Threadripper 3960x or 10 times faster than the 64-core Threadripper 3990x, lol.

That's what we call a constructive comment !
Now, if you can prove something or add something remotely interesting do it. For now, maths proves it's at least 5 times faster. But nobody has done a proper comparison on the same machine.

[Alex_M]

My comment was based on common sense. No need for math to know that it's unrealistic to expect a RTX 3090 graphics card that costs ~$1500 to be as fast as 640 CPU cores (10 x 64-core Threadrippers) that cost $40 000 for the CPUs alone. And that's excluding the rest of the hardware inside the 10 workstations that would amount to additional >$20 000 for a total of >$60 000 in hardware. So yea, I find it very hard to believe that a ~$1500 piece of hardware would be on par with hardware that costs >$60 000.
Anyways, I'm probably getting an RTX 3090 in a few days. I'll do a few tests vs the 32-core Threadripper 3970x on a few production interiors and exteriors so I'll let you know if I was wrong. I'm not expecting it to be more that 2-3 times faster though. We'll see.

[FranklinTre]

My comment was based on common sense. No need for math to know that it's unrealistic to expect an RTX 3090 that costs ~$1500 to be as fast as 640 CPU cores (10 x 64-core Threadrippers) that cost $40 000 for the CPUs alone. Anyways, I'm probably getting an RTX 3090 in a few days. I'll do a few tests vs the 32-core Threadripper 3970x on a few production interiors and exteriors and I'll let you know if I was wrong.

Well, my 300 € GPU is 5 times faster than my 1000 € CPU. On an pretty much equal generation, without counting the 400€ MB to plug it.
So, no, it's not about common sens. We're comparing 2 different technologies, not doing the same thing at all.

Still, maths is maths. But as you said, the only thing that matter is to compare the RTX 3090 with an equal price CPU right now.

[Alex_M]

Ok, so first results are in! I just got an RTX 3090. Soo.... yea. As I guessed, it is nowhere near 10-25 times faster than a 32/64-core CPU. At least not in the 1st test I did which is a production interior scene.

You can check the screenshot below showing the VFB history where you can see the render times. GPU finished in 6m 1.2s and CPU finished in 11m 4.3s which is not even 2 times faster (1.84 times to be exact).

Render settings were at default and Noise Threshold was set to 0.01 for both CPU and GPU. For the GPU rendering I used the faster RTX engine (not CUDA). Both the GPU and CPU were at stock (no overclock) at the time of the test. If I overclocked the CPU (which I normally do), the advantage of the RTX3090 would be even smaller. I'll probably do some more scenes when I have more free time. Like a production exterior for example.