[numerobis]

CPU vs. GPU

[AlexP]

while i can agree about limitations, it has to be said GPU engines on the market do their work, anyway. And a welldone mixed solution should take away nothing from Maxwell as is now, but just give an option for those who need it - correct me if i'm wrong.
Regarding speed, it would be needed to make some deeper experimentation but you can take what i said above: iRay renders on 8 GPUs a 5000 pixels scene which takes more than 1 day to get clean on my nine I7 CPUs on maxwell. 8 GPUs costs averagely the same than nine I7, consume much less and render averagely 5x faster.
Anyway i started another thread about Intel Xen Phy, these may change the rules and would be nice to know if NL will support these.

paolo

But iRay is fraction of Maxwell functionality. Try to turn off indirect reflection and refraction and see speed improvement.

[dmeyer]

CPU vs. GPU

I hope it doesn't take GPU render users that long to prepare their scene.

[PaZ]

But iRay is fraction of Maxwell functionality. Try to turn off indirect reflection and refraction and see speed improvement.

honestly i never dare to use caustics. I always use just direct and indirect light, and turnoff all other switches by default.... and it's way too slow for production anyway.
Unless you have a 30+ nodes renderfarm, but this is not our case.


Paolo

[AlexP]

Have you tested identical scene?

[PaZ]

do you mean in maxwell, with and without caustics ? Or in Maxwell and Iray ?

In first case yes, depending on scenes i can get indefinitely longer times.
In second case, no. I just saw demos on some classic virtual set, same complexity of most of mine. I know my average rendertime, and at same quality (iRay quality is extremely high), rendertime is a fraction with same money investment on hardware, as specified in a post before. But beside the demo, i know people working on GPU and i know their daily work.
Times wont compare at all. Quality yes - as said look at DeltaTracing works, entirely iRay based.
You can keep telling GPU is early and problematic, but results speak much more. This is my opinion, based on the fact i know more than a firm working on virtual set on GPU, and they produce much, much faster than me.
Btw i wont render a complex exterior on GPU, but that's another story. It's a dedicated tool, and for the area it covers you cant beat it or even come close.
And if iRay may have less accuracy of Maxwell (you wont notice it in most productons), there are other engines with all kind of accuracy - i.e. Octane, which is spectral-based as well.
This is not to say these engines are better than Maxwell. I love Maxwell. But production-wise, GPU road *IS* a concrete possibility for a large improvement, right now. All the rest are words.


Paolo

[archigrafix]

I feel GPU vs CPU is an obsolete debate... My last 3 archviz jobs (exteriors) where done with Octane. I like maxwell but when it comes to hard deadlines and real professionnal conditions speed is crucial! Working with Octane and not maxwell saved a max amount of time and stress.
I already tested Cycles in "hobbiiste conditions" and I feel it's got the best potential: CPU + GPU commuting from one to another freely.
Anyway a company like NL cant stay out from the GPU thing too long... They already are a little late...so I hope they are planning things in that direction soon

[PaZ]

that's why i'm asking.
I hope we can exclude NL will overlook at technology news and market trends, and will keep hardbeating CPU horse only.
Let's hope to hear something soon. We closed a big work (30+ 5000 pixel pics for bathroom catalog) done in maxwell, it was nice but toooooo slow with our 9 nodes and a heavy usage of external farms. We really need some drastic speed improvement in a short timeframe.

Paolo

[osuire]

whatever comes.. I will be soooo ready.. 2 x 690 GTX 4GB on order...

Well, it depends how long it takes to come, huh ?
By that time, those cards might already be re-purposed as paper-holders.

[dmeyer]

Just got a K20x into the office, in case any beta testing is needed. Hint hint