So now that we have a 32 core / 64 thread CPU confirmed and coming out in the next couple of months, is there a justification for investing in GPU rendering? Especially with the current GPU prices and availability and the fact that GPUs still don’t support all CPU features and are usually limited by their onboard RAM. I read that the 16 core 1950X is on par with GTX 1080ti in terms of rendering speed so the 32 core Threadripper 2 would be about twice as powerful as a GTX 1080ti with none of the downsides of GPU rendering. Is my reasoning correct?
So with this trend, any reason to go the GPU route if one gets 2-3 of these 32 core CPUs? On top of that, rumor has it next year AMD will be releasing a 64 core / 128 thread CPU! That’s insane if you ask me.
Your concerns are valid imho. But consider it this way - next nvidia gpu will be much faster then 1080ti, and in fact some of the more higher end cards are a lot faster and have way more ram 32gb in some cases today (though way more expensive). I think its going to be a sort of race now. I agree that the gpu is “promised” to give us a lot more then it currently does and in fact I think its overrated.
I personally can’t justify investing into a gpu farm because of some of the issues currently present and it has been like that in last 5 years. So I kind of made a choice to stay with cpu and so far its been proven to be reasonable.
You also have to factor in that on larger scenes no matter how fast your gpu is, some things just cannot be sped up, like initialization of the scene into memory, some of the processes are single threaded even on cpu, reading proxies from hdd, textures etc. So there may come a point where a scene renders in 5 min per frame on most powerful gpu setup where it takes 4 minutes to load and 1 minute to render. But if you have 10000 frame sequence no matter how you spin it, you would need a great number of those gpus individually to accomplish that animation even at 5 min per frame.
Lastly I think vray gpu is improving a lot, but any time I open and try to render something there is always an issue of some sort. Just too unpredictable at the moment imho, to do a tight deadline project and have so many unknown factors.
As always - if you start with V-Ray GPU from the start, it will be fine. Otherwise it is like switching engines during production.
Otherwise - I am pretty confident that the GPU rendering future is bright.
If you have scenes that you load slow and you send them over, we will fix it.
What’s most interesting to me is the direction major players are taking.
Renderman is heading on the hybrid route with XPU and maybe some deferred loading for textures. Still curious if it just conforms to GPU result while using CPU to emulate GPU codes.
Arnold is pushing really hard on making sure the GPU result is matching CPU. So using GPU for simpler scene for look-dev and bring CPU for the final frames if desired.
V-Ray is that the results between GPU and CPU won’t be the same, period. Depending on your scene, start with GPU or CPU and don’t switch.
I personally prefer GPU results matching CPU’s so people can have the option to switch or throw all the power in a PC for rendering without worrying some feature limitations or the results would look different.
RenderMan XPU is just like V-Ray GPU. They can run both on GPUs and CPUs, but result will never match RenderMan (in our case, V-Ray GPU will never match V-Ray).
Autodesk said that they will not support all the features on Arnold GPU. I am not sure how this works with “it will look the same”.
Blago the issue for me personally is not having the ability to load token variables which I use 99% of the time. So any scene I try with vray gpu immediately can’t load textures and that is it. I also use al surf a lot and have to wait until that is supported.
The impression I got after seeing their presentations is the goal is the results should perceptually match between GPU and CPU (not pixel-to-pixel accuracy). Will they ever get there with complex scenes? I don’t know. They only showed some screenshots of single frame of character assets that looks very close. Maybe it’s just the best scenario. I guess the pressure or the reasons behind that both (RMan and Arnold) are pushing really hard on matching GPU and CPU results is that lots of big vfx houses aren’t gonna use GPU farms for final frames yet, so it’s quite important for GPU-lookdeved assets look very close to CPU farm output, so there are less to worry about or troubleshoot down the stream.
Maybe Chaos is ahead of the game and already being there, done that, and knew how difficult it is to match between GPU and CPU results, hence suggesting to start the scene with either one and finish in that.
As I said already, RenderMan XPU (which is CPU & GPU render) is not trying to match RenderMan (which is CPU render).
It is trying to the extend where V-Ray GPU (which is CPU & GPU render) is matching V-Ray (which is CPU render).
Autodesk are saying that they will match the Arnold GPU to Arnold, but while at the same time they don’t plan to have all the features of Arnold in Arnold GPU. I am not completely sure how the matching works when they don’t support the same feature set.
Blago you are correct of course in saying that using gpu or cpu is like using different engines. I also think the logic of using gpu for preview in case of arnold and cpu for production is flawed: how can you reasonably predict the final render time / look on cpu when using gpu? Once the tokens are supported I would def love to try and run some full production shots with gpu only but for me it will take serious resources to build a gpu farm maybe in time
Maybe it’s all marketing trick to get users hyped up after all. Apparently, both emphasizing a match between GPU and CPU. Would be interesting if they ended up telling users to forget about matching the two, then we know how much we are ahead. I agree that the two will probably never match especially in terms of doing lighting for full production scenes. On the other hand, whether it’s flawed or not in terms of lookdev-ing smaller assets in GPU and be able to confidently assemble GPU-lookdeved assets in final lighting scene for CPU rendering, I am still very curious to see how it unfolds.
Leaked Cinebench results show AMD’s Threadripper 2 2990WX 32-core/64-thread CPU score ~5100 CB in cinebench R15, almost double the performance over last year’s Threadripper 1 1950X 16-core/32-thread CPU. Expected price - $1800 USD.
Benchmarks are out. Looks like in Threadripper 2990WX is a monster. Exactly 2 times faster than 1950X and 40-45% faster than the 18-core 7980XE (which costs the same) in Vray.
That’s because some run their CPU at default speeds (no overclocking), some run it with light overclocking, some run it with heavy overclocking and then there are people that have overclocked it pedal to the metal just enough so that it doesn’t explode in flames before the benchmark completes. The real results of 7980XE is 37 seconds and can be seen in the image I posted in this thread. That 27 seconds result was likely made with mad overclocking which was just stable enough for the short duration of the benchmark. As per the comments of the author of that 27 seconds score, he had overclocked his 7980XE to 4.8 Ghz on all cores and his CPU temp was 93°C. And that was after only 27 seconds of rendering! Most CPUs die at temps over 105°C and he was dangerously close to that threshold.
The problem with GPU rendering is ram. If your working on scenes that fit into GPU’s ram then it is fine otherwise it is a bust. 2080 series cards have no extra ram and we still don’t know if NVLINK allows shared ram on those cards like the quadro supports.
I guess the pressure or the reasons behind that both (RMan and Arnold) are pushing really hard on matching GPU and CPU results is that lots of big vfx houses aren’t gonna use GPU farms for final frames yet, so it’s quite important for GPU-lookdeved assets look very close to CPU farm output, so there are less to worry about or troubleshoot down the stream.
