Hardware Advice

[eid-o]

Hello everybody,

I am in the process of upgrading my workstation hardware, but I've come to hear yor expert advice whether it's better to spend more money on CPU or GPU. I know for years VRED has been CPU only, but the latest versions I hear allow you to render proper still shots (not viewport) with the GPU instead of hammering the CPU.

I am definitely going for AMD Threadripper, but there is a significant cost difference between the 32 core 3970X and the 64 core 3990X. I suppose the following question can be asked knowing if two GPUs (for example RTX3080 or RTX3090) will outperform a single CPU like these two in a still shot rendering - lets say 4k resolution. If rendering on the GPUs is now reliable in VRED and much faster compared to rendering on CPU I suppose it makes no sense to spend so much money for top of the line CPU. Has anyone made any tests of this kind to compare and share the results?

Another question related to the VR performance. Would you suggest installing a second graphic card makes a huge performance jump in VR? For example I am wondering if I will have a better performance having two RTX3080 10GB vs one RTX3090 24GB.

If the CPU vs GPU question above is in big favour for the GPUs...and two GPUs also make a lot more sense in VR then I suppose I can just hook two  of the latest graphic cards and extend the life of the current machine.

Many thanks in advance!

Best Regards,

M

[Christian_Garimberti]

Hi, waiting for an official answer i can share with you my experiences.

First of all keep in mind that officially only Quadro gpu's are officially supported by Vred, also if GTX are fully functioning.

Another thing: GPU rt in Vred is not only for stills. Also viewport rendering could be in Gpu rt.

I got 2x Quadro RTX 5000 and compared with the 2 x Xeon 3.6 6 core (so 24 threads) is 3 time faster.

Nothing like the threadripper, but if you think about the RTX 30 series i think i could say "GPU forever"!

The only thing is that you have only 10Gb vram in the 3080. I think that for rendering it is not enough...

So choose the right one...

For VR my advice is to have 2 GPU, but everything depends on the size and complexity of your scene. 

In conclusion my advice is to evaluate GPU instead of CPU.

Best

Chris

[eid-o]

Hi Chris,

Many thanks for your swift and useful reply, much appreciated!

It is indeed quite interesting to hear that the 2 Xeons with 12c/24t are 3 times slower than a pair of Quadro RTX5000. The point I made for rendering stills with GPU is because until recently that wasn't available neither in viewport nor for stills - good to hear that you can now use the power of GPUs when RT in viewport/production shots. This is why I am very very curious to find out what is the performance difference when you compare the best CPU on the market (the 3990X) against brand new GPUs such as the RTX 3000 series. For the price of the CPU alone I can get a duo of RTX3090...

For VR I wanted to find out whether two RTX3080 would outperform a single RTX3090. Some gamers are posting now that in gaming the difference is negligible, but we're not gamers and I am interested in VR for VRED. Scenes are sometimes complicated featuring full interior & exterior for several vehicles at a time. Adding hi-resolution shadows and HDRI lights starts to cripple the scene, hence I wouldn't hesitate to throw in another GPU.

Presuming VRED is GPU going forward in the future I suppose the following conclusion can be made: If 3970X/3990X are still hard to beat or close performance-wise to a pair of RTX3080, I'll get a 3970X with a pair of RTX3080, but if such powerful CPUs prove to be behind RTX3080/3090, then I'll keep the current 8c/16t i7 5960X and shove in two graphic cards. 

[Christian_Garimberti]

It is very difficult to say which option is the best without a real comparison.
The only problem you can find buying a 3080 it is the 10Gb Vram... Vred doesn't use shared memory, and load all the scene in the GPU ram. So I think also with 2 3080 you will have to stay in the 10Gb limit.
And in many cases, with mechanical stuff, I reached the 16Gb limit of my RTX5000.
But thinking about 2x3090 I think you will have many other problems...
Best
Chris

[eid-o]

Thanks Chris! That’s an interesting point I haven’t considered before - VRED being unable to use shared memory. So basically you’re saying that even if I have 2x 10GB cards VRED will only read the memory of one card?

What other problems you’re seeing with the RTX3090, difficult to source or the power consumption...or both 😂

Best Regards,

M

[Christian_Garimberti]

Yes, Vred (at the moment, but maybe in the future, who knows) can use only the Vram of 1 GPU.
What about the 3090? Room in the case could be a problem! It is a 3 slot card so, if you plan to put 2 inside your box... your motherboard should support it.
Best
Chris

[eid-o]

Thanks, your input has been greatly appreciated Chris!

i do have a full size tower with 1200w high-end PSU, so in theory the case will accommodate two of these beasts. I just need to double check the length. 
My motherboard is ASUS X99-E WS, which comes with 7 PCI-e slots. I have one used by the Intel PCI-e NVME SSD, hence 6 available. The intel i7 5960X supports 40 PCI-e lanes, the SSD requires 4 lanes if I am not mistaken while two RTX3090 will need 32 lanes - 36 lanes in total. If I decide to go down this route should be all good in theory! 

Still interested to hear if anyone has tested a high end Threadripper vs RTX3000 series.

[michael_nikelsky]

Hi,

as Chris already mentioned, the VRAM is the most important aspect when using GPU raytracing. Personally I would recommend 16GByte as the minimum so the 3090 is the way to go in my opinion if  you need  to stick to the Geforce lineup (which is not officially supported by us but will probably work just fine). If you ever plan to use 2 GPUs the 3090 is the only way to go at the moment anyways since it is the only Geforce 30xx supporting NVLink.

Memory sharing is not yet supported in VRED and you should not expect to have twice the memory once we support it since a lot of buffers need to be on each GPU to prevent massive performance loss (after all, NVLink is more than 10 times slower than VRAM).

Comparison to a Threadripper is  a bit difficult since CPU raytracing still supports some things that GPU raytracing does not yet support (Photonmapping, Spectral Raytracing, NURBS raytracing and some measured material types). So if you need those you have to use CPU raytracing and for that the Threadripper are the best way to go. However in terms of performance a Quadro RTX 6000 should be roughly twice as fast as a 64 Core Threadripper but we haven´t explicitly tested this setup. 

Kind regards

Michael

[michael_nikelsky]

Just forgot to answer the VR question: Using two GPUs in SLI for VR can give you up to nearly twice the performance in the best case. It all depends on what is actually the limiting factor and how your framerate on one GPU is since VSync and these stuff also influences the speedup you can get. As a more general answer I would say a speedup of 70-80% when adding a second GPU is more realistic.

That being said, the 3080 does not support NVLink so if SLI works (which I expect but am not sure of) the data has to be transfered over PCIE instead of NVLInk which is bit slower. Especially on high resolution HMDs this can become a serious bottleneck, especially with PCIE 3.0. I expect PCIE 4.0 to perform better but I don´t have any data on this yet and it will still be slower than a NVLink connection. 

So to sum it up: If you want to stick to the Geforce series you should invest the money for a 3090 instead of a 3080 since the additional memory and support for NVLink are huge advantages and they become even more relevant once we release memory sharing support for the GPU raytracing (which will definetely not be usable over PCIE alone and will therefore require a NVLink bridge).

Kind regards

Michael