United3dArtists Archive

Can Blender Cycles render produce finished quality renders using only the gpu? (pro, like architectural renderings)

What experience have you had with Cycles using modest graphics cards like my GeForce GT 610? (2 GB, 48 cuda cores)

I assume yes, but mainly with Nvidia CUDA cards only, I think OpenCL can be used, but still in development. You'd need a separate build for that.

But in essence, Cycles is a hybrid, it does both CPU and GPU calculations, obviously the GPU accelerated one being up to x4 faster.

Concerning your vidcard, for any GPU render engine, the vidram size is only useful for geometry caching, so the more polygons you have, the more ram you'd need. It won't affect speed. The clock and cores will. The more cores you have running at faster clock speeds will make it run faster. The wider the bus transfer rate is of the vidram will decrease scene load time, and the bit rate will generally affect how fast each core calculates together. So... you don't have many cores, so it will be a modestly slow render speed up, not much.. but with the two gigs, you can render quite large scenes without a problem.

Now, getting realism will depend on your talent, and the shaders you use, and work on the lighting a lot. But cycles can potentially do all of that.

Blender can make photo realistic ,just need knowlege of the nodes
And good textures as for gpu render the more cuda cores the better ,
I tested my gtx 780 and to my suprise my cpu was quicker
I7 3770

If 2300 cores isn't faster than 4 cpu cores then why graphics cards with gpu's? I guess that for video game level graphics the gpu's are awesome but for high realism (architectural renders etc.), much larger than screen size (print, cinema) the gpu is not so useful?

Well it's the clock speed... if you have 2300 cores only running at 800mhz, then it doesn't amount to much - maybe a bit more. Having 8 multithreaded cores running at 4.3ghz will crunch many bits faster. But sometimes you want the 2300 cores to calculate each little ray, and if you do it right it will be phenomenally faster.

To fully use your graphic card, it is better to use 256x256 tile in Blender, apparently they only calculate one tile at a time. But a multithreaded CPU will do many at a time, so better at 16x16 or 32x32 tile sizes.

Here is a good quick read to speed things up and make the most of your GPU or CPU.

Well for 4 cores seriously overclocked to 4.3GHz that is about 1.7 trillion core cycles per second. 2300 cores at 0.8GHz is about 1.8 trillion per second. (not really that simple of course but for a quick estimate of ability) That does mean similar performance of GPU vs. CPU as Pugman1 experienced. So again, with no performance boost even with a big card why do GPU cards exist? The page you (Drasie, thanks for that link!) gave notes a tremendous improvement by switching from cpu to gpu mode ... which would explain why people will pay hundreds of dollars for gpu cards. Of course, they buy the cards to play games not to produce pro illustrations or movies.

Again, thanks for the link to the speed tips. I'll just have to test and tinker when I get Blender. That is after getting work and real world stuff done so I'm not tempted to neglect that. That Cycles can produce pro quality renders is good with or without super speed. It still means no expensive renderer is needed and provides the ease of use (if that can ever apply to Blender) of having it integrated with Blender.

Thank you Draise ,maybe ill try the card again at the moment its in my gamer pc doing nothing
I gave up gaiming
Finis you should check out the blender forum there are some amazing architectural work
All done in blender the size of the pic is up to you how big you make it i stay. 1080 p

HI.Iv just done a test cpu against cuda. i installed blender on my pc with the nvidia card
tranfered my snapper file with all the textures set up the lightting so all was the same on both pcs
i7 3770k 18 seconds 1080p, gpu cuda 14 seconds 1080p .i dont think its worth me taking out my workstation
ati fire pro v7800 for 4 seconds more speed .blender is working on doing cycles rendering with ati cards .
at the moment they have a prob with the drivers , hope they get it sorted soon then i can
test ati against cuda ..

Well...... nearly 1/5th faster is... going to count if you are doing animation with 1 hour long renders per frame. Instead of 5 hours, you are only going to take 4 hours render - every 20 hours render, you've saved yourself a whole 4 hours render time on your machines along with 20% yield with the same time! It adds up.....

And if you renderfarm it over the network, you can have the other machine going at the normal rate, and you've more than doubled your yield, with the ability to still work on your other GPU machine doing compositing or modeling or something else (while it's rendering) without having to tie up your CPU and freeze work for the day, or week, or months - depending on the render load.

Sorry, just being ambitious, who would want to make an animation with 1 hour long frames?! Pixar!? Disney!?

So, big gpu cards are great for playing Call of Duty in HD at a million frames per second but for non-game, photo-real type rendering the gpu aspect of Cycles offers little advantage. Maybe some of the operations/calculations needed for "illustration" type renders aren't divisible into small similar operations for parallel processing.

Given that Pugman1's giant gpu made little difference then lesser ones will make no difference. I'll activate the gpu feature but consider Cycles a cpu renderer unless testing on my system reveals a significant boost. If it does, if a small card out performs a large one, then Cycles gpu is very sensitive to system configuration.

I'm certainly going use Blender as my main cg program since it is free and powerful but it is good to know that the free renderer, speedy or not, can do HQ stuff.

Maybe a bi processors i7 5960x with 8 core at 4gh?

marcel wrote:Maybe a bi processors i7 5960x with 8 core at 4gh?

I'll need a quad socket mobo ($4000) and some 15 core Xeon cpu's ($6,000 ea). Yeah! Sixty cores, $28,000. Overclocked please. My birthday is in March, hint, hint.

It is time for chip makers to stop the more cores thing and start making 128 or 256 bit processors. GPU makers should learn that there is a market other than gamers from HQ CG people (illustration, architectural, art) and make some cards with the cpu based render functions implemented in speedy specialized hardware plus the usual gpu multi-cores.

Nvidea quadro cards if you need power rendering ,cost way
to much though

hi all again ,just tested my gpu again against cpu ,render settings the same on both
1920. 1080 full100% samples set to final and 30 samples gpu tiles 30 30 cpu 16 16.
snapper model from zbrush 14 mb one texture map . eye glass shader and mixed shader glossy .
1 plane with ermision set to 20000 .you wil not belive this my cpu renderd this in 27 secs -
the gpu with every thing the same 57 secs .i tryed changing the tile size for the gpu no difference .
this makes me think that the cpu works together with th gpu when rendering. how can my i7 render quicker than cuda?
i7 ati fire pro 32 gig
i5 gtx 760 12ggig ------ not a 780

I think the best result for a picture bigger than the screen is sometime when the render is 100% cp. When both work together you Win time to calculate and you lost time with transfert data from cpu to gpu. The good solution depend of the hardware and the software. Render HD illustration is different than video game optimized for gpu with the resolution of the screen.

Thankyou Marcel ,i found out that low poly models render alot quicker than higher poly .
Low poly 10 secs 50 samples
High poly 27 secs 30 samples
I am am happy with my sytem it renders and works fine the way it is so im
not puting a nvidea card in just for a better render when its slower any way.