what would be a good 3d card for cycled?

Hi allAt my work place we have 15 dell precision workstations with core two duo and NVIDIA quadro 1700 cards.We don’t have funds to buy new PC’s but thought to get new cards to render faster with Cycles.Any one has a good suggestion for a pci e slot nvidua card for around 150 to 250 dollars? cuda of course has to work.ThanksClaas

Hi, best bang for bug was always the GTX x60Ti 2GB DDR5, if you can get older GTX 560Ti 2 GB take them.
GTX 500 series is about 10-20% faster as GTX 600 series for cycles.
For smaller budged get GTX 5/650Ti 2GB DDR5.

Cheers, mib.

Moved from “General Forums > Blender and CG Discussions” to “Support > Technical Support”

@cekuhnen - you continue to post support threads in the discussion forum. Please stop. We get plenty of traffic in the support forums and I’m sure your question will be answered quickly and adequately there (in fact, I’m pretty sure it’s been answered a number of times already).

Those Dell Precision T3400 workstations motherboard I think have PCI express v1 I think. The GTX 560 used PCI E v2 but as far as I know the cards are backwards compatible to PCI E v1.

So if I understand you right the GTX 560 is faster than the GTX 660 ?

Also does it matter if a GTX card is not from NVIDIA but from a 3rd party to run Cycles and Cuda with it?

Thats interesting.

So if I understand you right the GTX 560 is faster than the GTX 660 ?

Yes, NVIDIA produce the reference cards and 3rd party companies make, for example, different cooling systems (most better), overclocked cards, more memory and so forth.
The GPUs are the same.

Cheers, mib.

I’d really recommend to practice TeFoSe, the ancient and long forgotten ways of the forum search.
There are about 10 threads each week asking the exact same questions.

OT: @mib2 as I see it in your sig; Octane 1.1 is out, with glare, bloom and stuff :slight_smile:

@arexma, yeah man and the new RC with volumetrics,architectural glass and dirt is coming this week.</OT>

Sorry for hijacking your thread cekuhnen. :slight_smile:


If this lost art would be reliable and it would be good - you dont think I googled this before posting?

It might be worthy to actually make a sticky thread with this type of information - might make it easier to find it.

I kitted out my Hackintosh with GPUs when I built it last year and so I did extensive research on what’s best.

With all the searching through Blender forums etc I came to the conclusion that the GTX580 is the fastest you’ll get. You could go for the GTX590, but that’s basically two GTX580s in one package. I’d definitely go for the 590 because of this, if it weren’t for both GPUs making use of the same memory, which usually tops out at 3GB, leaving each GPU with only 1.5GB.

Try and hunt down some GTX580s with 3GB of ram, that’s my suggestion.

That’s because the new forum doesn’t allow tags anymore, which would make a search much more efficient.
Nothing personal, but answering the same question over and over again can be frustrating, especially if for the most times you clearly see that people did search, but were too lazy to actually read trough the threads the search spit out.


The relieable, proven, price to performance winner:GTX 560 TI.

Relieable, proven and fast in cycles: GTX 580 3GB version. Very fast, lot’s of memory.

Very good OVERALL price to performance but not as fast as it could/should be:
GTX 660 TI: Highly underestimated, this is a very fast card for gaming/display priced below 300 Euro/USD and there’s a 3GB version as well. Unfortunately the new CUDA toolkit for this card doesn’t work as fast as the one for the 500 series. The ASUS DiCU versions fan noise also is almost not noticable under full load, compared to other cards sounding like leafblowers.

Supposedly the top of the crop:
GTX 670/680: Fastest single GPU cards money can buy, however they suffer from a 10-15% lesser speed than their 500 series counterparts due to the CUDA toolkit. Might be fixed soon, in the future, or never. In games they’re faster.

Upcoming beast:
GeForce TITAN: rumored to come in february it’ll be a single card, the first real Kepler based GeForce with 6GB VRAM and rumors have it it’s over twice as fast as the 680, which would be a glorious performance gain. Estimated cost, 1000 Euro/USD straight.

In the end it boils down to:
Do you need the best GFLOPS to Euro rate?
Do you need the best Watt per GFLOPS rate?
Do you need the absolute maximum Speed available?
Do you exclusively need the card for CUDA?
What’s your budget?

Are you sure the performance between the 500 and 600 cards is down to the toolkit? When looking into this myself I had the opportunity to ask Thomas Dinges and he said it was down to the kepler architecture of the new cards, more cores, but those cores were less powerful than the cores in the 500 series cards. This wouldn’t be fixed by an update to the toolkit…


the search functionality in this forum is seriously flawed or not working well. Sometimes I find rather through google something here than using the forum search.

It to be honest might really be a good idea to create a sticky note with the information like the Cycles bench mark list.

This way then people will find the pure information much faster.


My bet is, it’s a combination of all.
Different architecture. Different toolkit. Might mean different way of coding stuff.
I also doubt Kepler will be considerably faster as Fermi with a “fix”, but I guess it’s possible to bring them on par at least - but better inquire with Thomas, Brecht :smiley:

Anyways, it might as well make sense that the Nvidia once again decided GeForce is for gaming and sacrificed CUDA performance for gaming performance - artificially, deliberately or not, who knows. :slight_smile:
Just like OpenGL had a serious performance drop introducing Fermi, the same might have happened to CUDA for Kepler.

True, I suppose us mere mortals will never know all the ins and outs as to why one is faster than the other.

Thank you for bringing up Titan though, looks like it’ll be a beast. My wallet isn’t thanking you so much though :’(