i am setting a new computer nowadays and i saw that the new blender cycles renderer is very powerful when gpu enabled. which graphic card do you suggest ? also which one is more stable with blender; ATI or Nvidia? Thank you.
Note: I am thinking of buying an asus gtx 580.
I’m using a GTX580 on Windows 7 and it makes Blender crash at least once per week (Error code: 7). I’d be interested to know if this is something that plagues others as well or if my system is the exception here.
Same question here. My employer will buy almost any card I ask for, so do I request:
1 x GTX560 Ti (384 CUDA cores)
2 x GTX560 Ti (can Blender actually use 2 cards to render twice as fast?)
1 x GTX580 (512 CUDA cores)
1 x GTX590 (2 x 512 CUDA core GPUs)
I understand the amount of RAM on the card is important, as the entire scene has to be loaded onto the card. So is a GTX580 with 3GB RAM a ‘safer’ choice than a GTX590, that splits its 3GB RAM between the 2 GPUs? Is 1.5 GB enough for ‘complex’ scenes anyway?
I’m left wondering if 2 x GTX560 Ti, each with 2GB RAM, might be the best choice?
I can see Cycles benchmarks people have done, but they doesn’t help me choose the best possible card. There must be thousands of Blender users who need to know the best current GPU for Cycles…
-Do you need to get a graphics card that works with Cycles?
-Can’t you wait a few months to see how the new Kepler cards by Nvidia perform?
-Do you have a big budget, but want to stay in the geforce price range?
If your answer to ALL these questions is yes, your best bet is 2 x GTX580s with 3GB of VRAM each. If you have the money, the PSU to handle it and spare room in your motherboard, you might even consider a third GTX580.
The latest Blender builds support multi-GPU in Cycles, but don’t scale very well. Typically, a second GPU will give you roughly a 30% increase in render speed. So in my opinion you’d be better off getting 1 x GTX580 3GB, instead of 2 x GTX560 Ti 2GB. It’s probably cheaper, performs better, has larger VRAM (needed for more complex scenes or higher quality textures) and will leave a free PCIe slot in your motherboard in case you want to add a new Kepler card in the future.
Now, if your budget constraints force you to choose between 1 x GTX580 3GB and 1 x GTX590 3GB, you will have to take this into account:
-Your GTX590 will work just as 2 x GTX580s with 1.5 GB each, severely limiting maximum complexity of your scene. GTX580 on the other hand, being a single GPU card doesn’t have to share those 3GB, so you’ll be able to fit larger scenes in it.
-At the moment, when it comes to performance in Cycles, 1 x GTX590 does not equal 2 x GTX580. At best it will perform like 1.3 x GTX580. This may improve in the future, however, but don’t expect such a great gap in performance right now. Besides, GPUs in the GTX590 are clocked at lower speeds than GTX580 and overclocking is severely discouraged unless you use extreme cooling (which will make the GTX590 an even more expensive option).
I hope this solves some of your doubts. Arexma, the local GPU expert might have something to add (in fact, I’m surprised he didn’t show up already). 
Thanks JuanJosé, that’s sort of confirmed what I thought. I didn’t know about Kepler…
I get a few PMs each day regarding hardware. I am glad to help, but after a while it’s getting boring to write something on the forum, repeat it in PM and then repeat it again in the forum because people don’t use the search.
To be fair, often enough people need “personal assistance” as the general answers to the posts just don’t match their situation.
But again, generally:
Seems only Tesla supports the unified memory mapping, where you can have multiple GPUs and and use each ones memory summed up.
A GeForce is much faster in single precision than a Tesla card due to it’s higher clocks. In double precision, which is for scientific calculations, not used in rendering, the Tesla is 8+ times faster, as the GeForce and Quadro are deliberately limited.
If you need 3GB VRAM for big scenes and got the money, get a GTX580 3G, preferably with a DirectCU (ASUS) or Frozr (MSI) cooling.
If 1.25GB is enough get a GTX560-TI 448 Cores, it has the best price to performance ratio. The 560 offers double the speed per Euro, however overall it’s slightly slower.
So it comes down to:
Do you need a frame to render in 30 seconds, or can you wait 35 seconds and rather have double the performance per Euro?
Do you need 3GB VRAM?
Do you need the card now, or do you want to wait for Kepler? (Where you might even have to wait till June)
asus doesnt have a 3gb 580 gtx card (nvidia). am i missing something?
I said “prefereably” with a DirectCU/TwinFrozr, should have been more clear on that:
ASUS aren’t available in 3GB, but they might be soon as all other board partners (EVGA, Gainward, Gigabyte, POV, Palit and Zotac) have 3GB cards.
MSI does too, even with a TwinFrozr.
You could always buy an aftermarket cooler if you’re up for some handywork. But the reference design cooling is loud and the cards is ~20°C hotter than with non-reference cooling.
If you intend to have long periods of rendering the reduces thermal stress can only benefit your chip.
arexma do you suggest Zotac? I have never used Zotac so i dont know how is the quality of the company. Right now i am stucked between asus gtx 580 directCU (which is 1.5gb) and zotac gtx 580 (which is 3gb) which one do you suggest me? thank you.
Like I said:
If you need more than 1.25GB VRAM you take a 3GB card.
If you don’t need it you get a GTX560 TI 448Cores.
Personally I think a GTX580 with 1.5GB is a waste of money. Poor amount of ram resulting in an overall bad price to performance ratio.
A GTX580 costs around 550 euro and has 1581 GFLOPS FMA
A GTX560 TI 448 costs around 270 euro and has 1311 GFLOPS FMA
The GTX580 1.5GB compared to the GTX560 TI 448 1.25GB:
Pay twice as much more for 17% more performance and 250 MB more VRAM.
In theoretical numbers, if a GTX580 renders 30s for a frame, the 560 needs 35s.
If you need to render 100.000 frames you might save a lot of time though.
I had 3 Zotac cards so far and despite them being a chinese board partner I never had any problems. They have quite a good reputation. But generally just google for “<product> <manufacturer> problem” 
arexma thank you for all the information. i learned a lot. my last question is i decided to buy a 3gb gtx 580 card. what do you think of this? http://www.zotac.com/index.php?page=shop.product_details&flypage=flypage_images-SRW.tpl&product_id=328&category_id=129&option=com_virtuemart&Itemid=100253&lang=en
thank you in advance.
The Zotac you choose is the cheapest but it has the crappy reference cooling.
If Zotac I’d take the AMP2! version. It’s manufacturer OC’d and has custom cooling, well worth the +30 Euro considering a decent aftermarket cooler costs ~50 Euro.
Depending on where you’re from you ain’t got much choise anyways:
http://geizhals.at/eu/?cat=gra16_512&xf=132_2048~1439_GTX+580&sort=p
To be honest, I didn’t even know that there was a 448 cores version. I thought they all had 384 cores. In the end it boils down to speed vs memory capacity, as usual.
Anyone considering a card with AMP2 should be aware that it will take up THREE PCIe slots, instead of the usual two PCIe slots. This is very important if you are planning for a multi-gpu setup.
Arexma: Thanks for your help on this! It all boils down to what a ‘large scene’ is, and I don’t know how a user finds out if their scene is going to need 1GB, 2GB or 3GB VRAM. This may be completely dumb, but if I switch to CPU rendering and watch the system RAM usage ramp up in Task Manager, will that give any sort of equivalent indication of how much VRAM my scene would need?
I can’t vouch for it, I do not know the underlaying sourcecode. If you run cycles in CPU mode though, the memory usage for the rendering should be equivalent to the one in GPU mode. But don’t rely on the task manager, there’s a statusline in the render window telling you how much memory it’s using.
I can tell you from my experience with Octane, that a 3.6m tris scene with a few small textures uses around 1.1GB VRAM.
So with a regular 1.25GB card you’ll hit the limit if you use textures.
The pure pixel information for a 4k HDRI map for instance is easily calculated:
40964069 pixel = 16.777.216 pixel.
A 32bit float uses 332b floats per pixel without alpha, being 96bit per pixel, = 1610612736 Bit = 201326592 Byte = 192 Megabyte.
A 8bit RGBA 4k image texture has 40964096(4*8) bit = 64 Megabyte.
So if you use high resolution texture maps your VRAM’s full in no time.