Graphicscard or CPU?

[PdZ]

Dear Chaosgroup,

Development of your GPU Renderer seems to get a lot better and it seems also a lot faster than CPU.

I just wanna know what is more wise to do?

Investing in Graphics Cards to render or into CPU?
Do you intend to abandon CPU in a while? Is there any focus?

Is it true that Graphicscard Rendering will be faster and faster in comparison to CPU in the Future?

Just wanna know where to put my money on it.

Thx in advance!
PdZ

[Muhammed_Hamed]

Is it true that Graphicscard Rendering will be faster and faster in comparison to CPU in the Future?

Hello!

I can only judge by the hardware available to us at the moment, the most expensive(consumer) CPU is Threadripper 3995wx or 3990wx at around 4000 Euros
This CPU scores 1200 in Vray's Cuda Benchmark
On the other hand an RTX 3090 scores 2000 points in the same benchmark, at 1600 Euros MSRP (when prices settle down)
Another important point here is expandability, any system can host 2 or 3x 3090s and then you can still expand that with an external box or Thunderbolt or PCIe. Even if you will not build this on your own, you can reach out to Maingear for example and order 4x or 7x GPU workstation
Conclusion is that GPU rendering is faster now and easier to expand. It is one big reason GPU rendering is very popular in the motion design community for example.

Do you intend to abandon CPU in a while? Is there any focus?

No, Vray CPU is an industry standard in VFX, Archvis and many industries. Usually you will see newer features come to the CPU engine first, then ported to GPU, the GPU development has its own team.

I just wanna know what is more wise to do?

Depends on your budget and what you do, are you using Vray GPU to render your current projects? or are you looking to get into GPU rendering?
Even if you invest in a stronger CPU, Vray GPU will be able to take advantage of that with Hybrid rendering.
The 5950X is a great piece of technology, even if you will be rendering on GPU. It is one of the best value CPUs at the moment, then you can pair it with a 3080Ti or a 3090. Then once you have a workflow, you can add one or 2 more GPUs to your setup easily
Vray's workflow is all about flexibility, If you tell me about your budget and the projects you do I will be able to help you better.

Best,
Muhammed

[PdZ]

Budget is not the question. Its more about being "future-proof". Dont wanna invest in outdated technology and later regreting the decision. CPUs are expensive. But you helped me a lot Muhammed. That was also my thinking.

[sirio76]

This CPU scores 1200 in Vray's Cuda Benchmark
On the other hand an RTX 3090 scores 2000 points in the same benchmark
Muhammed

To be fair you get that score while the CPU is emulating a CUDA device, an highly inefficient task

In some scene a top CPU will be faster, in other scene a top GPU will be faster, it really depends on what you are doing.
One of the advantage of the GPU is that you can use multiple cards (you can do that with CPU too but it won’t scale that well for render tasks) and you can easily replace them with newer generations keeping the same system, on the CPU usually socket are compatible for a only a couple generations.

[Muhammed_Hamed]

To be fair you get that score while the CPU is emulating a CUDA device, an highly inefficient task

I think this benchmark is quite reliable in this kind of comparison. Vray GPU got optimized to run on CPUs a while back and it performs much better now
Quite a few people use Vray GPU on local CPU farms with Hybrid, it gets the job done

In some scene a top CPU will be faster

If you have a scene where this CPU could beat a 3090 I would love to see..specially with Optix (the benchmark scene doesn't use Optix acceleration for this scene)

One of the advantage of the GPU is that you can use multiple cards

Ya, up to 8 and 10 GPUs per render node.. these do exist, while you can only have 2 CPUs in one machine with Windows 10 Pro

[seandunderdale]

If you can wait a year you might be able to get a 128 core / 256 thread CPU from AMD. I personally prefer CPU after working on GPU for 6 months. I found my speed gains were lost from bug fixing and re-rendering buggy renders. Also having to find workarounds for things that didnt work in GPU. This was last summer, so maybe things are better now, but I found my GPU experience to not be enjoyable. I also didnt like all the error codes vray threw out in GPU and add in which GPU driver is good or bad...made for stressful deadlines.

[sirio76]

I think this benchmark is quite reliable in this kind of comparison. Vray GPU got optimized to run on CPUs a while back and it performs much better now
Quite a few people use Vray GPU on local CPU farms with Hybrid, it gets the job done



If you have a scene where this CPU could beat a 3090 I would love to see..specially with Optix (the benchmark scene doesn't use Optix acceleration for this scene)


Ya, up to 8 and 10 GPUs per render node.. these do exist, while you can only have 2 CPUs in one machine with Windows 10 Pro

CPU do help when using hybrid rendering but only if you have a good processors, AMD CPU perform much better than Intel at this but both of them runs slower than using standard Vray CPU when running in GPU mode, this is not surprising since emulating a GPU is not a trivial task.

For the GPU VS CPU speed comparison this is the first example that comes to my mind: https://www.youtube.com/watch?v=myg-VbapLno you can see quite a few tests running a 3090 vs an i7 7820, given that a TR 3990x is about 5x faster than the CPU used in those tests it looks like the CPU beats the GPU in most cases beside the volumetrics (even in this test it would be really close to GPU though).

We have done some internal tests (of course comparing similar CPU/GPU) and looks like the GPU can be faster in some case but not by a large margin, sometime the CPU takes the lead as well so it really depends on what you are doing.
Beside that is true what seandunderdale says, while I do not see a 128core TR coming anytime soon (probably it will take a couple years), the CPU rendering seems more robust in production even if sometime the GPU can be faster, especially if you compare several cards versus a single CPU

[piotrus3333]

VRay GPU in 3ds max (and 3090) exclusively since february. perfectly fine engine but CPU is just more feature rich and robust. at least in 3ds max.

but if you are dedicated to Cinema then I would look at it from a different angle: other best render engine options for you are GPU based. so it all depends on your definition of “future proof”.

[Muhammed_Hamed]

but both of them runs slower than using standard Vray CPU when running in GPU mode

I don't know exactly what you mean by this, the standard Vray CPU and Vray GPU are 2 different engines. You don't run standard Vray CPU in GPU mode, the terminology is wrong
Hybrid rendering runs the Cuda code on CPU, which gives identical results to GPU rendering.. so the standard Vray CPU is not involved here

The point of comparing a 3090 vs AMD 3990wx using the Cuda engine, is that the results and noise will be identical. Which makes for a fair test in my view
It is true that Standard Vray CPU could be faster and it will give different results, because it is a different engine.. not because "the CPU is emulating a CUDA device, an highly inefficient task"
Read about Hybrid rendering here,
https://www.chaosgroup.com/blog/unde...brid-rendering

It’s important to note that the V-Ray Hybrid (GPU–CPU CUDA) renderer is not the same as the V-Ray Production (CPU) renderer, and the two engines will continue to remain separate.

Even if this very expensive CPU (4000 Euros) could be faster than a 3090 (1400 Euros MSRP) ,how many CPUs can you fit in one system to expand your setup talking about "future proofing"?
Here is my setup a few years back, there are quite a few consumer boards that can fit up to 7 GPUs

[sirio76]

Which makes for a fair test in my view

While I agree with you that emulating a CUDA device is the only way to get 100% identical results, emulating a GPU is not the ideal way to squeeze all the speed from a CPU you can take an M1 SoC and run X86 code on it but running native code will be faster, sometime way faster and this is using a system highly optimized to do so.
About the price, well MSRP is one thing, real world price is a totally different story, I just look on Newegg and the cheapest 3090 was twice as much, it will come down to MSRP eventually, as as soon as the next crypto mining bubble explodes (who knows when, then wait for the next bubble).
About future proofing you are right, a GPU based system can last longer, you can surely install newer cards easily (even several cards if the MOBO support them), eventually after a few years you will be still be forced to upgrade the whole system since newer GPU may require larger PSU, triple slot, more cooling, more PCIe lanes to get full bandwidth ecc.