Gpu rendering

[3di]

I've just rendered 350 frames at 1080p using fstorm in 3 hours on one gtx1070.  The same animation rendered on 4 computers took 40 hours with corona :(  corona's like a wife I adore, and fstorm the busty  seductress I'm trying to look away from.  Come on corona, get that boob job!

[Christa Noel]

haha I like that funny statement :D
btw, fstorm 3hr 350frame from only 1 gtx1070 vs 4computers 40hr corona, that's an interesting battle. what cpu(s) did you use?
would you like to share the video here?

[3di]

:) i7-7700k 32 gb ddr4 , i7930 overclocked to 4ghz 24 gb ddr3, and two quad core xeons 16gb ddr3 and 8gb ddr3.  Can't share the animation unfortunately, it's an old job still under NDA.

Not the most amazing setup, but I'd probably have to spend around 6k to match the £400 1070.

[Christa Noel]

afaik they corona will employ the power of gpu only for .. kind of features accelerations not for rendering engine.
example: Accelerate bloom glare process in VFB post pro.
but who knows? everything seems more possible after the corona joins the chaosGroup

[3di]

Yeah I hope so, I'd hate to leave corona behind after using it for so long.  Can't argue with a £5500 saving in hardware though.....and that's if I only want to match one 1070!

[Juraj]

I also rate you highly for creative writing :- )

Personally I haven't seen any speed benefits in F-Storm from those who actually use it to render complex interiors like our favourite swedish boy duo.

That's not to detract from you gpu request, I personally hope that could eventually be viable upgrade at some time, after all, I am looking at those 16-24gb GPUs to become mainstream next year :- ).
GPU accelerated Framebuffer with Denosing/Bloom&Glare/etc.. will be fantastic start.

[3di]

Yeah, and with nvlink hopefully coming to GeForce cards next year 2 x 8gb cards will equal 16gb video ram instead of 8gb.   AMD vega 64 cards can also treat system ram or ssd as gpu ram at the moment too thanks to their new infinite fabric technology.

I must admit I haven't tested with an interior scene yet, but heavy refraction and outdoor scenes were MASSIVELY faster.  When I let that bad boy rip it was a similar feeling to  watching a game load almost instantly from a sega  master system cartridge after years of watching the ocean loading screen on the trusty c64 :)

[Juraj]

Exterior were always blazing, 6 years ago I remember them being almost instant with first alpha of Octane on old GTX 285 with 2GB of memory. Hence the tradition from which the "rotating Car on youtube" Meme came :- )

And interiors took endless, up to 100times more to clean-up ('clean' is the essential part). Seems like that didn't change much since then.

[3di]

Damn it, I'm going to have to pause my box set binge now to go and compare interior scenes :)

[3di]

I don't normally do interiors so the quality isn't the best, just an old scene I downloaded from a free site ages ago and slightly modified/set up for corona.  It has a forest pack rug, and mainly transluscent materials for shades/curtains with a bit of scattering if I remember correctly.

I've set it to render at 1920x1080p with a noise level limit of 1.5% and rendered in corona on 4 machines (dr), then i've changed the renderer to Fstorm, hit the convert scene button, changed the hdri bitmap gamma to 2.2 and enabled direct lighting and then hit "run rt mode" using 1x gtx1070.....oh I also tweaked the lamp shade material a little as it was too see through in Fstorm initially (I may have gone a little too far though as i've robbed the fstorm render of quite a bit of the yellow light.

On corona, I waited until the time remaining had stabilised and was ticking down consistently, but on Fstorm I let it run all the way through to 1.5% noise due to it's time remaining not being as accurate as corona's as it constantly decreases as it renders.

Results attached.  Fstorm was still suprisingly favourable still coming in not quite as quick as the outdoor scenes, but still approx 8 times faster.  So I definitely think corona should seriously consider having a change of heart on their cpu only stance...with their expertise and team size i'm sure they could match what one guy has been able to do on his own with fstorm.

[cecofuli]

For a serious comparative, you should use the same light setup. Also, don't use any converter, but convert by hand with the better setup.

If you want to try, this is my old test, converted by me, from FStorm file to Corona 1.3 (yes, I should try with the new Corona v1.7) .
n
--- LINK---

If you look the 18 minutes FStorm test (Titan X + GTX980Ti) and my test  with very old i7 970 six core, (15 minutes)  Corona wins.
Why? Because with less rendering time and with very old CPU I have more or less the same quality or, even, better quality than an expensive dual GPU rendering.
Of course, we should try with the new version of Fstorm and Corona v1.7. Also the Fstorm original rendering ( I found it in the FStorm forum) was in JPG and not in PNG.

But, I'm very curious to see a test with two rig with the same price. For example, a new 1000 GPU (two GTX 1070) with new 1000 CPU (ThreadRipper 1950X) with the last version of both softwares.

PS: remember that in Corona v1.3 there weren't any bloom or glare.

[andrewdtejeda]

For a serious comparative, you should use the same light setup. Also, don't use any converter, but convert by hand with the better setup.

If you want to try, this is my old test, converted by me, from FStorm file to Corona 1.3 (yes, I should try with the new Coroa v1.7) .

--- LINK---

If you look the 18 minutes FStorm test (Titan X + GTX980Ti) and my test  with very old 7 970 six core, (15 minutes)  Corona wins.
Why? Because with less rendering time and with very old CPU I have more or less the same quality or, even, better quality than an expensive dual GPU rendering.
Of course, we should try with the new version of Fstorm and Corona v1.7. Also the Fstorm original rendering ( I found it in the FStorm forum) was in JPG and not in PNG.

But, I'm very curious to see a test with two rig with the same price. For example, a new 700 GPU with new 700 CPUm with the last version of both softwares.

PS: remember that in Corona v1.3 there weren't any bloom or glare.

If you can, I would really like to see a comparison with the newest version of fstorm and the newest version of corona. My studio may be making a render engine change soon and this would be very helpful in the decision making process.

Thank you!

[cecofuli]

You can do by yourself!
I'm my original thread, you can download the FStom .max file and the Corona .max file.

--- LINK---

https://drive.google.com/open?id=0B_i7C0g1XqRMN1lodFNUb1d5REk

I'm not FStom user, sorry. I used the FStom DEMO version, one year ago.

If you will do the test, please, post the result ;-)

[3di]

Hi Cecofuli, your quite right.  I hadnt noticed but the hdri, although the same one, it was automatically flipped 180 degrees during the conversion process so the light is streaming into the room from a different direction.  I'll retry and see how the result varies.  I'm using 1.1.10v by the way.

[andrewdtejeda]

You can do by yourself!
I'm my original thread, you can download the FStom .max file and the Corona .max file.

--- LINK---

https://drive.google.com/open?id=0B_i7C0g1XqRMN1lodFNUb1d5REk

I'm not FStom user, sorry. I used the FStom DEMO version, one year ago.

If you will do the test, please, post the result ;-)

Ahh, I had hoped you were.  Neither are we, and we don't have access to the newest version of it. Maybe someday, once it's available, I can take a whack at it.

[3di]

flipped the enironment of the fstorm version to match the Corona version, and it made quite a big difference to fstorm render times.  Around 3 times slower than before.

[3di]

You can do by yourself!
I'm my original thread, you can download the FStom .max file and the Corona .max file.

--- LINK---

https://drive.google.com/open?id=0B_i7C0g1XqRMN1lodFNUb1d5REk

I'm not FStom user, sorry. I used the FStom DEMO version, one year ago.

If you will do the test, please, post the result ;-)

Ahh, I had hoped you were.  Neither are we, and we don't have access to the newest version of it. Maybe someday, once it's available, I can take a whack at it.

I've downloaded the scene provided by cecofuli, and rendered in corona 1.6.3 and also fstorm 1.1.10.  I didnt change any settings in the corona version apart from changing it from a 15 minute render to a 1.5% noise level instead.  In corona I rendered it on one i7-7700k running at stock speeds, 32 gb ddr4 ram.  In Fstorm I used a gtx1070 gpu (also shared by windows). To get it rendering in Fstorm I just changed the renderer, hit convert from the render tools panel, pumped up the lights considerably and then tweaked the tone mapping so the images roughly matched.  Results attached.

Corona 2 hours and 10 minutes rendered to 1.5% noise.
Fstorm 25 minutes also rendered to 1.5% noise.

[Christa Noel]

O_o shit! 2 housr vs 25 minutes ?! what a very interesting comparison.
Can anyone here provide the most fair scene for renderEngine comparison purpose? honestly, I'm doubting the current scene somehow.
does fStorm support Radeon GPU? and where can I download the Fstorm demo version (stupidly I cant find the download link)? :D

[maru]

I really do not want to get involved in this, so just one note for now:
1,5% noise limit is probably a different thing in Corona than in fstorm. In Corona it is an overkill, as values between 3 and 5 should usually produce good results. You can see that the Corona rendering is very smooth, while in the fstorm example there is still some noise visible (e.g. at the top of the image).
Other than that, 2 hours is definitely too long to get good quality on a 7700k, so it was either rendered for too long, or there are some issues in the scene, which affect render time.

I think a bit more fair comparison would be like this:
-make a scene render to acceptable quality in fstorm in the lowest time you can
-then try to beat this score in Corona
(or vice versa)

[Christa Noel]

..1,5% noise limit is probably a different thing in Corona than in fstorm..

noted

[cecofuli]

Of corse that it's different! =) A hahah Just too on the white furniture or on the ceiling.

Corona is almost ZERO noise, FStorm has a lot of noise (please, post PNG and not JPG)
For a good comparison, you should post the same rendering time, not the same noise level ;-)

[maru]

For a good comparison, you should post the same rendering time

...which is dependent on render settings, bias...

[3di]

Hi guys, i'll re-render to times instead of noise.  I'm pretty new to fstorm so i'm not yet sure how to optimise the render settings for that engine yet.  Corona looks pretty well optimised though.  Settings are are all default more or less, portals in place, albedo all good.

@Maru.  Do you fancy checking the scene to see if anything can be optimised?  I'd be interested to see if we can speed corona up any.  I might even learn something! :)  The link is above in cecofuli's post.

[3di]

O_o shit! 2 housr vs 25 minutes ?! what a very interesting comparison.
Can anyone here provide the most fair scene for renderEngine comparison purpose? honestly, I'm doubting the current scene somehow.
does fStorm support Radeon GPU? and where can I download the Fstorm demo version (stupidly I cant find the download link)? :D

Hi Christa, it's nvidia only I beleive.  For any questions regarding the engine though, such as availability etc then your probably best messaging the page admin over on their facebook page:

https://www.facebook.com/groups/FStormGroup/

[romullus]

Hi guys, i'll re-render to times instead of noise.  I'm pretty new to fstorm so i'm not yet sure how to optimise the render settings for that engine yet.  Corona looks pretty well optimised though.  Settings are are all default more or less, portals in place, albedo all good.

No doubt, eventually they will steer you to the settings where Fstorm will start rendering slower than Corona :]]]

[3di]

Hi guys, i'll re-render to times instead of noise.  I'm pretty new to fstorm so i'm not yet sure how to optimise the render settings for that engine yet.  Corona looks pretty well optimised though.  Settings are are all default more or less, portals in place, albedo all good.

No doubt, eventually they will steer you to the settings where Fstorm will start rendering slower than Corona :]]]

I can't believe that.  Maru has been very helpful.  He even just private messaged me to let me know that an older geforce fx series graphics card may help me clear up the noise faster in fstorm :p  Just kidding Maru :)

[cecofuli]

Of course that rendering time will depend of the hardware.
If we use a Geforce 560 vs Ryzen it won't be a realistic match =)
And, an opposite side, it isn't realistic to use a i7 920 vs 1080 GTX .

but, in this situation a 7700k (300 euro) vs 1070GTX (400 euro) can be a good compromise.

But, it's absolutely better to use rendering time instead % of noise.

[3di]

yep, just doing a rendering time test instead.  I may have been doing fstorm an injustice.  Apparently I should have been using the latest nvidia drivers for a 20% boost in render times.

[cecofuli]

And, from Corona "Injustice" =) , you should use a AMD Ryzen 7 1800 (400 euro) and you will have 70% of boost in Corona ;-)

PS: Maybe you have to reset Corona setting. I see some strange jagged edges on the window. Maybe some old, hide setting inside my old file.. I don't know...

[3di]

ah ok, cool i'll try resetting the settings.

[3di]

Just had chance to test them again in the way Maru and Cecofuli suggested.  This time I rendered fstorm until it looked to be an acceptible noise level and then rendered with corona to the same time of 25 mins.  There's a bit of a speed increase in Fstorm with the latest graphics drivers.  Corona also performed much better than last time after resetting the render settings (not sure why, the only thing that changed was max sample intensity raising from 10 to 25 and the filter changing from 1.5 parabolic to 2 tent.  Same hardware was used as last time.

[Juraj]

So we moved from "it's 10 time faster" to "it's the same" ;- ) ? As confirmed by the actual f-storm pros and not car rotators.

What picks my attention:
F-Storm has perfect AA. Does it clamp the result ? If you save the result as 32bit .exr and take it into Photoshop, (or anywhere else), is there full dynamic range ?
F-Storm really has quite nicer lighting on the ceiling. Is this due to different material only (more reflective) ?

[3di]

Hi Juraj

I didn't change the materials at all, just hit the convert button and turned up the lights a bit, then tweaked the tone mapping.  Just checked and all values are the same for both the corona and fstorm ceiling material.

The other questions, i'm too new to the engine to be able to answer.  I'd recommend you ask over on their facebook page.

As for the noise....I thought there was a noticeable difference still in Fstorm's favour, but perhaps I've just spent a little too much time today being an over analytical geek :D

[daniel.reutersward]

Hello,

I had to jump in since I use FStorm in my daily work and have been for the past 1,5 years.

Hardware wise my 2x Xeon 2697v3 cost very similar to my 4x Titan X (Pascal), at least here in Sweden.

Overall regarding render times when I have tested new Corona versions (which I do from time to time) I feel I get similar render times compared to FStorm.

Perhaps I need download the scene and see what result I get :)

@Juraj: Yes, you have full dynamic range if you save an image as 32-bit exr, at least when I opened it up in Affinity (don´t have PS anymore) :)

/Daniel

[Juraj]

Daniel, when you open the images as 32bit .exr, is the AA still so perfect ? If so...that's quite some feature.

I really want to get into what makes the light behave tiny bit more directional in F-Storm (which is very nice), we've been analyzing this with Dubcat and it's quite interesting that it really is there.
But I can't put my finger if it's due to AA, different HDRi sampling, or shaders. Esp. since you have diffuse "roughness", where lower values can make materials more plastic and more suspectible to light. It's always very noticeable on ceiling, and below the window.

[daniel.reutersward]

Far from perfect in the two images I tested with! Unless you go really dark.

Yes, I´ve been talking to him too a bit. Very interesting to see his tests :)

[burnin]

one of last year's tests, 10min with Luxrender (OCL) @ luxrender forums.
Has bit better AA.

ps
Corona is doing fine on CPU :)

[3di]

Wow, so many good renderers out there at the moment! 

I'd say corona's doing more than fine on cpu, it's an incredible engine, not to mention it's ever increasing range of excellent features, and stellar customer service.

My main reason for the request was that looking to the future, gpu's are cheaper to have multiple of,  due to many motherboards being able to house up to four,  they're cheaper to replace/upgrade as pcie is less likely to change, and even if it does it's generally backward compatible unlike ever changing cpu sockets which often require a new build, and they hold their value well due to a massive gaming market.   With all that in mind, if you then consider that even a single consumer gpu can either roughly match a similarly priced cpu in some rendering scenarios and vastly out gun it in other scene types...then it just seems like a good direction to consider.

[3di]

Ok, final post from me.  I obviously hadn't been using the engine for long enough on the previous experiments. 

After learning a little more about kernel settings, removing the portal brought over from corona, removing refraction from the windows, and overclocking my 1070 to 2025mhz I was able to get the scene rendered to roughly the same level as before in just over 4 minutes, with a fraction more noise I was able to render it in just over 2 mins.

Peace out.

[Ondra]

already tracked here, moving to resolved:
https://forum.corona-renderer.com/index.php?topic=96.0