I’m hoping to build a new system for myself sometime in the near future. Does anybody have advice when it comes to selecting a GPU?
My choices are the GTX 660 Ti, GTX 670, and GTX 680. I’m hoping to game some but cycles/CUDA performance is important too. Given that, which of these cards gives the best bang for the buck?
Thanks!
Moved thread to Technical Support
Thanks for the move, wasn’t quite sure where this thread should go.
If you are set on the 6xx series then I would recommend getting a GTX 670 with 4GB VRAM (unless you can afford the 680 with 4GB). More VRAM helps with bigger scenes. If you want the absolute fastest render times then look for a GTX 580 with 3GB. The 5xx series are still around 15% quicker at rendering in cycles then the 6xx are.
Hope that helps,
Jon
Thanks for the info Jon, I will be upgrading within 6 months hopefully.
Thanks for the advice!
Yeah, I’m set on the 6xx for a few reasons. First, the 6xx series has fully optimized out-of-the-box support under OS X, which is a must as this is a hackintosh build. The 5xx series also works, but was never optimized as Apple didn’t use any 5xx series cards in their machines, instead opting for AMD offerings. The other is gaming performance since the 6xx has a good leg up on the 5xx series there in most benchmarks I’ve seen. Kepler cards are also more power efficient and cooler, which makes keeping the system cool a lot less of a headache.
How noticeable is that 15%? It doesn’t seem like that much of a sacrifice.
Personally, I would give up the 15% for the space since I often go over 3GB for a scene and find myself having to do a lot of optimizing. But if you are doing animations you may gain hours/ days from the boost.
Best of luck with the Hackintosh.
if you are planning on using CUDA then a nVidia GeForce GTX690 will be the best as it has 3072 CUDA cores which is over double that of the GeForce GTX 680 so yeah that would be the best but the best bang for the buck NO the 690 is like $1,000 so that works out to around $30 per core and CUDA cores only work with stuff using CUDA. In that the 690 also is basically 2 680s on one board 2 sets of GDDR5 2 GPUs etc. thus the very high price though.
Out of what you said and going by prices today on www.newegg.com here is a table of how they stack up
Model | CUDA cores | Price
Geforce GTX 660 Ti | 1344 | $269.99 - $349.99
GeForce GTX 670 | 1344 | $359.99 - $459.99
GeForce GTX 680 | 1536 | $439.99 - $799.99
GeForce GTX 690 | 3072 | $999.99 and $1,049.99
So using CUDA cores / Price we get:
Model | Min Price per CUDA core
Price worked out as Min Price / cores
Geforce GTX 660 Ti | $0.20
GeForce GTX 670 | $0.27
GeForce GTX 680 | $0.26
GeForce GTX 690 | $0.33
The amount of CUDA cores matters as more will decrease how much time it takes to run a CUDA job and Blender has tons of CUDA jobs especially if you are working with multiple objects, textures, high res, high sample, and small tiles as CUDA works better with bigger tiles.
Now with that in mind you also have to look at which one is how much are you using CUDA, are you gaming on that machine, and how much are you willing to spend on just the graphics card. As i will suggest not getting the GTX670 but the 660Ti or the 680 and another note will be all of them are made for PCI-e 3.0 slots. You will want one with at least 1GB of Graphics RAM as holding textures and textures some times are very big file sizes though i do not know how big res you have yours usually better to more then 2GB as anything over 2GB is just a marketing gimmick so the 4GB graphics RAM on the 690 might be used as its only 2GB per GPU and it has 2 GPUs compared to all the rest having only 1 GPU.
Do you want speed or speed per performance as speed per performance i will go with the 660Ti for just pure speed i will go with the 690. In what i just said both cost as much as my entire computer put together but of course mine wasn’t built for graphic design.
A good deal of misinformation there.
Cuda core count is NOT the most important thing to look for in a CUDA card. The 5xx series Fermi architecture is severely more capable than the 6xx series Kepler when it comes to computational capability. For example: the GTX 580 (512 CUDA cores) still outperforms a higher clocked GTX 680 (1536 CUDA cores) by at least 15% in compute speed. Also, memory bandwidth - sure, both have GDDR 5 but that on the 5xx series is 384bit while the 6xx is only 256bit.
CUDA cards do NOT only work with programs that use CUDA. Nvidia cards are also OpenCL compliant.
A GTX 690 is indeed 2 GTX 680s on one card. However, to get them both on there they are slightly underclocked. See the difference in something like the Mars II versus a GTX 590. Furthermore, the 4GB VRAM on the 690 only works as 2GB in GPU compute applications.
PCI-e 3 hardware is fully backwords compatible with PCI-e 2 and shows minimal gains in practical performance.
If you can, wait to see if nvidia’s next card will fix what’s wrong with the 6 series; 3gig is a bit low imo; even 4gig isn’t great, but the next big thing (whatever it happens to be) cant be far off. 
It’s a case of balancing your needs for now, with your frustration at having missed out, if you’d waited, on something awesome.
Agreed. I was planning to build a new render machine with a bunch of 680s but figured that it should be just a few more months until the 7 series launches - which should finally be at least as quick at rendering as the 5 but with more RAM and cooler operation.
More VRAM helps with bigger scenes
i thought cycles resolved that with the tiling technique, or maybe NOT?
with tiling you could render any scene you want with limited RAM (at least what i think).
The tiling helps but still doesn’t solve things (unless that changed with 2.65 - which I still have not installed yet).