BIGGEST BLENDER CYCLES GRAPHICS CARD FAQ's

Been trying to figure out the whole gpu product issues and figured this might be helpful information for those trying to figure out what to buy for Cycles.

//Which card to get CUDA or OPEN CL?
So far CUDA is supported the best so any Nvidia or other manufacturer CUDA based card should work. Besides, Nvidia cards support open CL also. However, look at the specs to determine compatibility.

//On the card, how many processors should I get?
It depends on how fast you want to render. Here is a list (towards the bottom of the page) that lists several Nvidia cards and what they contain.
In essence for Cycles- you’re looking for the 1. Amount of processors 2. Core Clock Speed 3. Amount of RAM (Which I don’t think is listed in the link)

//How do those factors make a difference?

  1. The more processors, the faster it will render
  2. The faster the core clock, the faster it will render
  3. The larger the RAM, the larger the scenes you can fit for realtime rendering (otherwise, if you make a change, it will have to reload portions or most of the scene to update versus immediate changes)

//Is there a major difference in speed between- for example- a GTX 570 and GTX 580?
From experience in using both cards- not much. But yes, the GTX 580 obviously is slightly faster.

//There are more than one version of the same model. What do the versions mean?
For example, EVGA makes a GTX580, GTX580 SC, GTX580 Classified, GTX 580 Hydro Copper, and a GTX580 3072.*
The differences to look for regardless of the cute names is simple. Processor speed and onboard RAM. In the case of the Hydros- a different build for water-cooling enabled systems.
GTX580- 772Mhz/ 1536 GB RAM
GTX580 SC (Super Clocked)-797Mhz/ 1536 GB RAM
GTX580 Classified- 855Mhz/ 1536 GB RAM
GTX580 3072- 772Mhz/ 3072 GB RAM
GTX580 Classified 3072- 855Mhz/ 3072 GB RAM

*The memory clock speed will also vary
So with any card, look for those changes between the models. Technically, you should be able to overclock any card, and most graphics cards include utilities for that, but the overclocked (Higher Mhz) versions have stronger components and better cooling specifically designed for running the processor harder than stock specs.

//Does buying an overclocked version mean that the processor is running harder?
Yes. Many cards will drain more energy and get hotter because of this. You will get faster clock speeds however.

//What is SLI?
NVIDIA® SLI™ technology is a revolutionary platform innovation that allows you to intelligently scale graphics performance by combining multiple NVIDIA graphics solutions in an SLI-Certified motherboard. -Nvidia Translation: You can link more processors for faster rendering. See HERE for an example.

//Can I can get greater speed by buying two or more cards and linking them using SLI?
Yes. You will multiply the speed additively. Sort of… see arexma’s full explanation below…

//Can I also get more RAM that way for larger scenes?
No. SLI mode is like giving 10 artists a portion of the painting to work on. But RAM does not additively increase (e.g. 1GB RAM + 1GB RAM= 1GB RAM for the processors to use). SO if you buy two cards at 1GB and you have a 500MB scene, they will BOTH load the 500MB scene and work on different portions, then give it back to the display. Thus, no matter how many cards you have- it only gets faster (parallel computing)- not more RAM. Reference from Nvidia can be found here…

//I want more than 3GB of RAM. If I buy an Nvidia Tesla with 6GB, will I actually have 6GB?
Yes.

//So why are the Nvidia Teslas sooo expensive and they may have fewer processors and can be the same speed as the GTX’s?
According to EVGA, Tesla’s have double precision and other professional features that may not be needed for the beginner or standard level 3D artist professional. (Reference)

//I have an old computer, will a standard NvidiaGTX model work in my computer?
It should, but obviously its newer technology and tested and designed on modern equipment, so its always better to have an updated system. Also keep in mind that since the graphics card’s speeds are faster than an old motherboard’s bandwidth, the system is slowing down the card.

//The GTX590’s are dual GPU’s. For my old computer, will that work for me?
Hard to say. Personally, I had an old motherboard and the GTX580 worked fine, whereas the GTX590 gave me errors and did not boot. The representative told me that the dual gpu’s are new technology and so they are not sure what they DON’T work on. So as mentioned, to be safe, have a fairly recent motherboard.

//I heard that GTX590’s have two gpu’s but will only recognize the single processor RAM amount- is that true?
I have not found any indication that this is false. SO, if you buy a dual gpu GTX590 with 3GB of RAM, it should act as if it has 1.5GB of RAM

Hope this helps someone in knowing what to purchase since there is a lot of technology going around and most of it is obscure to the average person simply looking to upgrade or use Cycles with the gpu. Anything to add would be much appreciated.

R3sili3nt

Thank you arexma! Thanks for adding the additional info!

Anytime. Maybe change your thread title to something more general as I partially already took the thread OT :slight_smile:

Some more:

// Is CUDA or OpenCL faster?

Several universities took up the job to painfully and scientifically compare the two. Overall it´s not much, but CUDA is faster.
OpenCL has the advantage though that the CPU can contribute as OpenCL device - over the obvious hardare independency advantage :wink:

// CUDA all good and nice, but how does Nvidia compare to AMD under OpenCL?

There´s not much benchmarking going on for that. The best thing is to compare the results of LuxMark, Luxrenders benchmark tool. I guess it´ll become the FOSS to CineBench at one point :wink:
http://www.luxrender.net/wiki/LuxMark_Results
Well not much is an overstatement, there´s virtually no benchmarking with results going on for OpenCL, who knows why.

Some additional Information provided by Blender.org:

http://wiki.blender.org/index.php/Doc:2.6/Manual/Render/Cycles/GPU_Rendering

As a side note, show this thread to lazy AMD driver developers to motivate them (maybe not to developers, but marketing team or even their boss?). First official release of most used free software package with modern render engine is NVIDIA only. What a shame.

Isn’t Quadro designed for workplace and Geforce for gaming?

Yes- but its like saying a race car versus a tuned, fast standard car. They both can have speed- just built for different purposes.

Quadros are designed to run 24/7

@R3sili3nt and arexma

Thanks for all that usefull information.

If I buy a mainboard with six “PCI-E x16 Gen 3.0”-slots.
Can I put a GTX580 in each slot or will it be limited by some technical restrictions ?

If I can put 6 GTX580’s on one mainboard:
A GTX580 uses about 244 W, I guess I need some specific poweradapater for 6x244 W ?

How will Cycles render an animation ?
I guess each graphiccard will render a frame (card 1 will render frame 1, card 2 frame 2 and so on…) ?
Does that mean that the first 6 frames will be rendered parallel and should be (theoreticaly) finished at the same time ?

Will my mainboard be overheaded by so many graphiccards ?

Kind regards
Alain

No, Cycles prepare each frame using CPU, apply all transformations, modificators, scripts, and generate final triangle meshes data structures, with materials. Make some accelatation data (BVH) as well. Then all that passsed to GPU hardware, and a lot of parallel GPU threads (kernels) started. So, basically it render 1 frame at a time, on all possible GPU hardware. You can cheat it starting many parallel BLender instances, skipping some frames, good old hack. But I doubt it will be faster, only more memory pressure.

@Storm

It means Cycles uses ALL six graphiccards for the first frame (a gtx580 has 512 cores, it means that there are 6x 512 cores avaible)?
Does this also mean it will be rendered six times faster than on only one graphiccard ?

What about the VRAM ? if one frame renders on all avaible graphiccards only one graphiccards VRAM will be filled with the scene and the remaining VRAM of the other cards is empty ?

Kind regards
Alain

It means Cycles uses ALL six graphiccards for the first frame (a gtx580 has 512 cores, it means that there are 6x 512 cores avaible)?

Yes

Does this also mean it will be rendered six times faster than on only one graphiccard ?

In theory it should but in practice it does not. So for each card that you put on the motherboard you get diminishing returns.
There is this crazy bastard that put 8 GTX 580s together for rendering in octane. I doubt you are interested in that amount of power :slight_smile:

What about the VRAM ? if one frame renders on all avaible graphiccards only one graphiccards VRAM will be filled with the scene and the remaining VRAM of the other cards is empty ?

Arexma explained this well enough above. Read carefully. If you do not understand google some more.
Vram is shared in all the cards so if you have 8 GTX 580s in there ( 7 of them having 3GB vram and 1 has 1.5 Vram they all get 1.5 Vram limit )

@reC

Thanks for your explanations and patience :wink:

I know about that VRAM-Limit when you have more than one card.
I just wanted to be sure that is a HUGE WASTE of VRAM because you if have 18 GB (6x 3GB of the GTX580) and it’s only possible to use 3 GB (=17%).

It would be MUTCH MORE efficient if you could render each frame on one graphiccard.

Edit:
Sorry, my mistake: Rendering each frame on a another graphiccard makes no difference

Isn’t there a smart solution for using more than one graphiccard ?

Kind regards
Alain

I just wanted to be sure that is a HUGE WASTE of VRAM because you if have 18 GB (6x 3GB of the GTX580) and it’s only possible to use 3 GB (=17%).

That’s just it. You DON’T have 18 GB V-Ram from 6 580 cards. You cannot even say that. Not even in theory you can’t because it doesn’t work like in CPU ram where it gets “added”.

Isn’t there a smart solution for using more than one graphiccard ?

Perhaps in the future we might see a technology like that but for now this is what we got and we have to make due.
If you’re interested in finding the most cutting edge technology you are free to keep yourself informed using Google search. That’s how most of us do it anyway.

Remember, PCI-E bus have very small memory bandwidth compared to internal GPU caches, many orders of magnitude. Even if it possible to access memory of 1 card from another using some page interrupt trickery, it will be 10000 times slower, no one need that.

Hi everyone, I’ m planning to buy a laptop with an nvidia gpu: wich one will be more worthy between gt 540m and gt 630m?
Thanks in advance for a response

To be more precise I’ m a bit confused about compatibility: will also gt 630m be capable of gpu render with Cycles?
And also, will it make a difference to get a 2GB version instead of 1GB?
I hope that someone have ever tested it…

Tobe more precise I’ m a lot confused about compatibility, could the gt 630m also be compatible with gpu render incycles?

Thanks again

Finally!
It is supposed to be my answer, so series 6xxm should work with cycles gpu accelleration
:slight_smile:

Hi everbody, I have one question which is not answered here: How much VRAM will a “normal” scene use? It is woth to spend 70$ more for 4GB instead of 2GB?
I hope somebody will read this question and can answer it, even though this thread is very old.

I guess you mean RAM (=RAM on your motherboard) instead of VRAM (=Video RAM on your graphiccard) ?
Because 2 GB more VRAM is alot more expensive than only 70$ :wink:
The maximum amount of VRAM for the recommended graphiccards GTX580 is 3GB.

For “normal” architecture visualization renderings I never ever had a VRAM problem (I only have 1.5 GB VRAM) since Tile Rendering Option is implemented in Blener 2.64.

Kind regards
Alain