Hi,
could you advice a good but cheap graphic card to use confortably with blender ?
(maybe cude technologie) ?? 
Well it depends on what cheap you mean.
The best mid-range card for cycles with cuda is the GTX 560 ti or no ti.
I am planning to buy the gtx 660 but I guess 6 series cards perform less then 5 series in cycles.
thank you for your answer.
what do you think about that ?
or that one ?
Well honestly I wouldn’t suggest buying the gt series for blender or rendering.
If you are tight on budget then i would suggest buying a gtx 550 or gtx 650 atleast. In my opinion they are the cheapest one suitable enough for blender.
And if you could add more to your budget, the best option would be the gtx 560 2gb (based on people review) for blender and cycles.
ok no problem.
what do you think about :
http://cgi.ebay.fr/Inno3D-GTX550Ti-3072MB-mHDMI-PCI-E-GTX-550-3072-MB-3GB-SDDR3-Dual-DVI-mini-HDMI-/230894099310?pt=DE_Elektronik_Computer_Computer_Graphikkarten&hash=item35c25c5b6e#ht_3956wt_1140
its on my budget…
i just wanna know if its really most performed than my HD radeon 5850 : for blender and use a Geforce for video editing (premiere cs5) ? i thought that cuda should be an answer of my research in performances…
don’t you ?
GTX 650 Ti 2GB is pretty cheap and renders Mike Pan BMW benchmark scene in 1:51
It’s also very quiet and has low power requirements.
I believe it’s not that bad for a card of it’s price.
in that case the best card might be the gtx 560 ti or gtx 650 ti 2gb as they both perform almost equally.
I wouldnt suggest buying gt 630 for blender.
I think the best card for you in you budget will be the gtx 650 ti 2gb.
For general use, amd cards seem to be better.
For rendering Nvidia is currently the only option for rendering via CUDA in cycles.
Other render engines support AMD acceleration; at least luxrender does.
Go for more memory as opposed to more cores. Slightly slower rendering on your gfx card is better than issues due to insufficient memory which results in crashing.
Which build of blender?
My stock-clocked 920 (4.5ish years old) rendered that scene in 3m 11s on windows with MingW (build 51011); Linux would have rendered it quicker still - sadly I didn’t use linux back then for rendering so no clue what it would have been. So the difference wouldn’t be that great and no memory issues using a CPU.
So really think what you want, before spending cash on something that may be less than ideal soon. Bare in mind that CUDA is getting increasing problematic - at least that’s my understanding - as time goes on with cycles.
Someone - for example - wanting a really beefy system could use 4 580s or 680s; they’d only have 3 or 4 gig max of memory to play with. The cost of those four cards wouldn’t be much different than a xeon or the equivalent amd dual setup. The dual amd setup may actually cost less.
I’ve thought a time or two recently of buying a 680 4gig; £500 quid would get me thisawesome, and awesomely expensive board. Of course there are cheaper ones available and I’d still need to shell out for a couple of xeons /cry. If I go Xeon in my upgrade, that is likely what I’ll do - unless Intel’s Haswell release makes me rethink.
Official 2.65a from blender.org site.
Same time on windows and linux. GPU render time is not affected much by compiler or OS version.
In windows, the GPU is 40 seconds faster than a 3930K 3.2Ghz (2:33)
In linux, the GPU is 14 seconds slower than the same CPU (1:37)
Not bad for a GPU with a price of a third of the CPU.
I agree with you that if possible, more ram an more power is better. I wanted a 4GB 680 but was out of my budget and I decided to spend more on the CPU than on GPU.
Agree, but no memory issues is awesome imo.
I’d be curious what a recent mingw64 build did with your 3930k. 
I’m also really curious to see what, if any, options the Xeon Phi will provide.
I can’t test mingw64 on windows anymore, I replaced it with linux as my only OS. I guess the speed would be similar to linux version. Mingw64 is gcc like in linux, so the compiler optimizations should be very close with some variation due to the underlying OS and threads implementation.
I noticed I wrote the wrong time for linux, it’s 1:37 (1:36 with r53622 from buildbot), not 1:51. 1:51 is the time for GPU.
Thanks to everyone for your participation in that topic 
Inno3D GTX550Ti mHDMI 3072 PCI-E GTX 550 3072 Mo Dual DVI SDDR3 3 Go mini HDMI
and
4GB Grafikkarte PCIe Nvidia Geforce GT 630 PCI Express 4096MB HDMI DVI VGA
are the two cards on my budget, do you think that i could have good performances with one of these ?
Appelation for three graphics cards:
Gt, gts, Gtx, in ascending higher performance. So gtx is essential.
Very important memory is crucial, DDR3 is discouraged, opted for GDDR5.
Do not buy graphics card on ebay.
Two reliable manufacturers, Asus and EVGA.
Here is a shop or service and prices are correct (be careful they are in France)
http://www.grosbill.com/index.php#gt-home-header
And the card that I bought Saturday
http://www.grosbill.com/4-asus_nvidia_geforce_gtx_660_dcii_o_-169237-informatique-nvidia
Hi, I second Deloince JP don´t buy this 3 GB gtx 550 with DDR3 VRam.
I know from a german blender forum it is very slow for Cycles rendering iirc 3 minutes for bmw benchmark.
My EVGA gtx 550 2 GB DDR5 VRam need 1.50 minutes but is a few dollar pricier.
Cheers, mib.
thank you so much for your advices
I had a GTX580 and it rendered the BMW benchmark in 1m13s. I upgraded to a GTX680 4GB Classified edition and it actually took longer to render. So I’m not sure what to do, bring it back or keep it.