For the context, I come from Unreal background where a lot of client work has been done in Unreal, Lumen. For 50 fps 4-6K resolution LED screens.

I am looking into switching part of the rendering pipeline to Karma XPU, but I expect I will not get as fast rendering speeds as with Unreal Engine.

Would this setup be good enough for production rendering in Karma XPU?

In total 3 computers with:

3x RTX 4090
2x RTX 5090
2x 9950x3D
1x 9980x Threadripper

Each computer equipped with 128 GB ram.

Thx for any advice.

My current system

AMD Ryzen 9 5950x (16 cores)
128GB of DDR4 @3200
Nvidia RTX 4070 TI Super (16GB VRAM)
I am looking into buying a RTX PRO 6000 Blackwell with 96GB of VRAM at the moment

Not a bad system, but the 16GB of VRAM is a bit short at time, but the 32GB of you RTX 5090 should help with simulations and render, and yes, render times will be much slower than with Unreal, by a magnitude, We are still waiting for Karma real-time, assuming that it is still in the works!

Don’t waste money on an RTX Pro unless you really need that much VRAM. It won’t be any faster than a 5090 or even a 4090. And for the price of one RTX Pro, you could probably get two 5090s.

Unless you specifically need 64 GB of VRAM for your renders, it’s basically a waste of money.

As for the number of GPUs, it always depends on what you’re rendering and how fast you need the results. For my work, 6×4090 across three machines is more than enough — but it never hurts to have more, as long as the return justifies the investment.

Mirko Jankovic
Don’t waste money on an RTX Pro unless you really need that much VRAM. It won’t be any faster than a 5090 or even a 4090. And for the price of one RTX Pro, you could probably get two 5090s.

6×4090 across three machines is more than enough

Thank you for this info. Now I have some starting point.

Since you are on Linux did you have any troubles handling Nvidia drivers for Karma renders? I understand that Nvidia driver stack on consumer GPUs on Linux is inferior as opposed to what is available for server GPUs like A6000 or workstation RTX Pro.

For server GPUs Nvidia offers Linux drivers with better VRAM management, shared memory allocation, and VRAM overspilling to system RAM.

Something they don't offer on consumer desktop GPUs on Linux (like 4090 or 5090).

No problems with nvidia on linux, actualy in some cases it renders twice as fast as on windows. No exaggeration. Depending on project I can see speed up in XPU rendering, from 10% to even twice as fast.

Again unless you need more then 32GB of ram for your renders you wont see any speed up in rendering at all, it can even be slower.
It mostly comes to CUDA cores, 5090 got 21K and A6000 only 10K.

Mirko Jankovic
It mostly comes to CUDA cores, 5090 got 21K and A6000 only 10K.

Ok I understand. So unless you hit VRAM limits 5090 or 4090 is fine.

Have you been in situation when you exhausted your VRAM? like with large environment renders? Or you pay close attention to optimization?