Author Topic: Corona GPU  (Read 41714 times)

2015-06-19, 13:40:26
Reply #15

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 8922
  • Turning coffee to features since 2009
    • View Profile
There seems to be a large number of paying Corona users (myself included) that want GPU + CPU rendering to happen. It would be interesting to see some performance numbers on GPU vs CPU rendering times for various setups.
https://embree.github.io/papers/2014-Siggraph-Embree.pdf :D
Rendering is magic.
Private scene uploader | How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2015-06-19, 14:23:26
Reply #16

dfcorona

  • Active Users
  • **
  • Posts: 198
    • View Profile
I've seen you post this before Ondra when this topic arises, but what is this suppose to prove.  In the real world results cpu embree cannot hold a candle to the speed of gpu.

Xeon E5-2699 v3 Octadeca-core (18 Core) $4,764.19 = around 775 GigaFLOPS

Nvidia GTX TITAN X $999 = over 7 TeraFLOPS

Real world test Scene I've tried. Titan X vs 3930K Overclocked

Vray CPU 11min 32sec - Without DOF

Vray GPU 1min25sec - With DOF

Octane full Pathtracing on interior scene 7min.

we can try real world tests, and not just papers.

I think you've done a great job with Corona Ondra, I am not trying to take anything away from you, but to look at now and the future of Gpu vs Cpu. I cannot see not implementing Gpu like everyone else.  What will Cpu's be like in 2016, maybe 10% increase in speed if we are lucky.  The Pascal video card in 2016 is suppose to be 10 TIMES faster than Titan X.

« Last Edit: 2015-06-19, 14:44:53 by dfcorona »

2015-06-19, 18:40:31
Reply #17

DanGrover

  • Active Users
  • **
  • Posts: 21
    • View Profile
I think you've done a great job with Corona Ondra, I am not trying to take anything away from you, but to look at now and the future of Gpu vs Cpu. I cannot see not implementing Gpu like everyone else.  What will Cpu's be like in 2016, maybe 10% increase in speed if we are lucky.  The Pascal video card in 2016 is suppose to be 10 TIMES faster than Titan X.

I don't mean to sound like a meany, but do you really think all this has passed Ondra by? I mean, I assume he has a lot more knowledge about the under-the-hood specifics of render engines than any of us, and he's said he's constantly observing changes in the GPU industry (so it's not a blanket "no"), and if there are engines out there that offer massive speed increases - as you seem to suggest there are - then why write a new render engine? Apparantly the market is already providing you with a render engine, no? If VRay RT is 10x faster than the CPU version (and I love Corona, but it's obviously not 10x faster than VRay generally) then... don't you already have your wish granted?

2015-06-19, 19:42:30
Reply #18

agentdark45

  • Active Users
  • **
  • Posts: 562
    • View Profile

Real world test Scene I've tried. Titan X vs 3930K Overclocked

Vray CPU 11min 32sec - Without DOF

Vray GPU 1min25sec - With DOF

Holy crap is this true?? A ten time render speed increase is not something to ignore! Can you post up some images please? This is making me reconsider Vray...

I don't mean to sound like a meany, but do you really think all this has passed Ondra by? I mean, I assume he has a lot more knowledge about the under-the-hood specifics of render engines than any of us, and he's said he's constantly observing changes in the GPU industry (so it's not a blanket "no"), and if there are engines out there that offer massive speed increases - as you seem to suggest there are - then why write a new render engine? Apparantly the market is already providing you with a render engine, no? If VRay RT is 10x faster than the CPU version (and I love Corona, but it's obviously not 10x faster than VRay generally) then... don't you already have your wish granted?

I think this is a moot point. We are all in search of the best overall rendering software, why not try to make Corona as good as it can possibly be? Most of us are here and have decided to part with our money because we find Corona better / more desirable than other rendering software as a whole. However the speed of a renderer is definitely a HUGE part of what makes it "good". Are you saying you wouldn't care if Corona could be 10x as fast as it is currently? 10 hours vs 1 hour - there is simply no arguing with that. Remember the early Maxwell days - quality, but day long renders! This was a big turn off for most people.
Vray who?

2015-06-19, 20:30:05
Reply #19

burnin

  • Active Users
  • **
  • Posts: 1055
    • View Profile
Would be nice to go Hybrid way with what you already have & OCL... it's coming along just nicely. Tests show it working on Lux, Indigo & Thea. Also VRay started with OCL but switched to CUDA later on... hmmhm.
You've seen new AMD Fury? http://www.forbes.com/sites/jasonevangelho/2015/06/18/amd-radeon-fury-x-benchmarks-full-specs-new-fiji-graphics-card-beats-nvidias-980-ti/
It is getting interesting...

2015-06-19, 21:05:56
Reply #20

dfcorona

  • Active Users
  • **
  • Posts: 198
    • View Profile
I think you've done a great job with Corona Ondra, I am not trying to take anything away from you, but to look at now and the future of Gpu vs Cpu. I cannot see not implementing Gpu like everyone else.  What will Cpu's be like in 2016, maybe 10% increase in speed if we are lucky.  The Pascal video card in 2016 is suppose to be 10 TIMES faster than Titan X.

I don't mean to sound like a meany, but do you really think all this has passed Ondra by? I mean, I assume he has a lot more knowledge about the under-the-hood specifics of render engines than any of us, and he's said he's constantly observing changes in the GPU industry (so it's not a blanket "no"), and if there are engines out there that offer massive speed increases - as you seem to suggest there are - then why write a new render engine? Apparantly the market is already providing you with a render engine, no? If VRay RT is 10x faster than the CPU version (and I love Corona, but it's obviously not 10x faster than VRay generally) then... don't you already have your wish granted?

Here is the answer to your question, Corona is brand new, built from the ground up with the latest and greatest.  Look at the speed and quiality of it's GI engine, plus it's simplicity, shader system, even though it takes longer to clean final render it's still very fast for what it is.  Imagine all of that running on the power of GPU or assisted on the GPU.  I believe Ondra has built a fantastic CPU renderer, and would love to see it be better. That's all...... and why not.

2015-06-20, 15:59:18
Reply #21

cecofuli

  • Active Users
  • **
  • Posts: 1571
    • View Profile
    • www.francescolegrenzi.com
VRay CPU vs VRay GPU

1 TEST: 19x
1 TEST: 23x
1 TEST: 10x
1 TEST: 11x
1 TEST: 19x
1 TEST: 17x

I'm sure that, soon or later, Corona will be CPU+GPU (in some way).
Right now, it's better to be concentrate to add a new features on PCU version.

2015-06-20, 18:40:40
Reply #22

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 8922
  • Turning coffee to features since 2009
    • View Profile
We are all in search of the best overall rendering software, why not try to make Corona as good as it can possibly be?

This is actually what we are trying to do, but hey, I am programming renderers only for 6 years, and those green-black marketing slides from hardware companies have to be at least a bit correct, right?
Rendering is magic.
Private scene uploader | How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2015-06-21, 01:21:30
Reply #23

bnzs

  • Active Users
  • **
  • Posts: 16
    • View Profile
And I was a little thrown over to the fan:
CPU is sad. AMD not give us in last 3 ears not one good cpu. Intel in terms of good jump on 2600k then on 4770k (or 3930k someone - that) and now 5820k. And this is in shot words 5 ears of life. And they don give us posibility to use several cpu's on one MB, even xeon - 2 cpu for MB is this standard version is above all more expensive so that it is easier to use DR. All other options is one - DR but for DR for each cpu we must have another MB, another 32gb ram, another SSD+HDD, another PSU.
BUT for 4 very fast gpus for example need only two powerful PSU that it.

Good example
and
conform this http://render.otoy.com/octanebench/results.php?sort_by=avg&singleGPU=1 8xGTX 580
it is = 4x980 Ti and they easy can stay in one case in sense PSU (2x1000W) and mobo - 4 PCIe its easy, and 980ti have 6gb vram. And again - look on speed in video in octane and vray.
I just hope that ondra in a secret from us develop corona on GPU. Even maxwell
think about gpu :)

And some more videos for GPU+corona vote (i hope Ondra already see they all) - in description render time\per frame
and if they give to users even 75% of what they promise here
http://home.otoy.com/otoy-unveils-octanerender-3-worlds-best-gpu-renderer/ it's strange that Ondra not interested in gpu and again I just hope that ondra in a secret from us develop corona on GPU.

2015-06-21, 02:00:35
Reply #24

dfcorona

  • Active Users
  • **
  • Posts: 198
    • View Profile
We are all in search of the best overall rendering software, why not try to make Corona as good as it can possibly be?

This is actually what we are trying to do, but hey, I am programming renderers only for 6 years, and those green-black marketing slides from hardware companies have to be at least a bit correct, right?

Ondra, why are you getting defensive? You are taking this the wrong way, we think your brilliant programmer for what you have done with corona. Why do you keep bringing up marketing slides or papers on the subject to dissuade us when we are telling you with real world tests the gpu is much faster for interactive setting up a scene to final rendering.  So far every scene I have worked on testing with cpu and gpu, there is no question that gpu is X times faster.  What will happen to this great renderer you created when next year Pasal comes out and all the other renderers have full GPU support for all features.  No matter how great corona is, people will move on if the render time goes from 4hrs, to just a few minutes.  and not even the fact of cost, gpus are cheap compared to the optional cpus.

Here what is great about GPU cost.

1 Titan X is X times faster than the best CPU.  Now for just $999 more you can add another Titan X reducing render times by half again.  And we all know most systems will house 4 GPUs.  And with pci risers you can put up to 8x GPUs.

compare that to the cost of one system that can barley compete with one Titan X. then multiply the cost of each new complete system you have to buy.

2015-06-21, 08:12:35
Reply #25

fobus

  • Active Users
  • **
  • Posts: 363
    • View Profile
We're using Corona for 2+ years for production and all this time we're had no doubs that it was right choice. Rendertimes was a bit higher than Vray, but simplicity using of Corona complitely outweights it. But now we have a big project with huge number of images to render at high resolution and render times becomes a bottleneck in pipeline. As we tested Corona and even IrMapped Vray has nothing to do with it. But our test revealed solution in VrayRT GPU as it is a 3x times quiker for the same money (42x i7-4790k roughly the same speed as 8x GTX 980Ti in two PCs). And much more compact and energy efficiet too (40+ PCs versus only 2 PCs).

Of course it is neede completely new code for GPU, new shaders and so on. I know that even all standart maps has to be completely redone to be rendered on GPU. But it has done in Vray and partially in Octane. GPU is really fast and it has a bright future I think. Even one or two yeras ago we can't even imagine that graphics card under $1000 can have 12Gb of memory. Since Titan X and introduction of Pascal with up to 32Gb of memory it will be possibe to to really big projects with rendering on GPU without strong memory limitation (that was a main reason to avoid GPUs).

2015-06-21, 14:02:50
Reply #26

Juraj

  • Active Users
  • **
  • Posts: 4209
    • View Profile
    • studio website

Real world test Scene I've tried. Titan X vs 3930K Overclocked

Vray CPU 11min 32sec - Without DOF

Vray GPU 1min25sec - With DOF

Holy crap is this true?? A ten time render speed increase is not something to ignore! Can you post up some images please? This is making me reconsider Vray...


Would like to see the comparison, because the old one Cecofuli posted is deceiving. It's not Vray vs VrayRT GPU, but VrayRT (CPU) vs VrayRT (GPU), where the engine, is pure BF/BF in both cases.
It just recently received LC algorithm as well, making even VrayRT(CPU) fast, but it's not fast as regular Vray algorithms.

But given I now have Titan-X as well, I can give it some try.
talcikdemovicova.com  Website and blog
be.net/jurajtalcik   Our studio Behance portfolio
Instagram   Our studio Instagram, managed by Veronika

2015-06-21, 20:04:37
Reply #27

fobus

  • Active Users
  • **
  • Posts: 363
    • View Profile
Just a small test to see GI speed. Lighting from sun and sky only. No portals.

VrayRT GPU (GTX 580)
Bruteforce+Bruteforce = 32 min
Bruteforce+Lightcache = 10 min

Corona 1.0 (i7-3930k@4.1GHz)
PT+UHDC = 27 min

As you can see VrayRT GPU BF+LC already at 10 minutes is clean enough.
Of course it's not a complex test but one of tests

2015-06-21, 20:13:37
Reply #28

fobus

  • Active Users
  • **
  • Posts: 363
    • View Profile
One more test at scene from Vray's forum converted to Corona correctly. (of course you can see some differences but I do not think that they have a huge impact in overall speed)

This test was made to know how many time it will take to render same scene as we needed for our project. We see that GTX 580 in VrayRT GPU has done it in 1/3th of time needed to render same scene in Corona on i7-3930k and i7-2600k.

Render times are:
VrayRT - 12 min
Corona - 12 and 29 min

2015-06-21, 20:19:01
Reply #29

dfcorona

  • Active Users
  • **
  • Posts: 198
    • View Profile
Fobus, thanks for the comparisons, can you post the tea pots scene, I have same processor as you, but I have a Titan X which is much faster than a 580. I'll run it and post my times for comparison.