Good video card/computer for cycles

I am thinking about building a computer specifically for blender. i think the most important part is going to be the video card. As far as cycles, what characteristics of a video card are most important for good performance?
I know that higher Nvidia card series does not necessarily mean better cycles performance. 5xx series works better with cycles then the 6xx series. So what should i look for to make a good purchase?

is this a good place to look?:

I don’t think this chart is just video card performance for gaming. I think it measures “direct computing” abilities. Not OpenCL, but direct compute. Apparently there is a difference. I am not well versed in these matters. So would the above link be a good place to look to find out how certain cards will perform with cycles?

Any info on the matter would be appriciated
Thanks all!


  • DirectCompute is Microsoft’s API - not supported because it’s not cross-platform.
  • OpenCL is Khronos (OpenGL folks)'s API - half-supported, can’t run Cycles on AMD cards because of driver issues
  • CUDA is Nvidia’s API - supported

So your choice is pretty much limited to Nvidia cards, the GTX series. A more reliable benchmark is This one

I’d look for cards with more memory on them - you won’t be able to render very complex scenes on the GPU if you have 1GB or lower.

What`s your budget for GPU?

Stay far away from 1- 1,25 -1,5 GB cards…, most of 500 series.

Thanks for the info. it looks like the gtx 580 is the way to go. even more so then the titan!

around 500$ (American)

I don’t understand

What about 2 cards (SLI)? What are the rules with that? Do they both have to be the same cards? If they are the same cards, will this double the performance? Is SLI complex and difficult to set up and use?

Two cards of course work well put make sure you have the power supply for that in the PC.

I think you dont need SLI but the cards should be the same - or best same RAM because the card with the least amount will be the amx cap. Having a 2 GB and a 1 GB card only will allow you to render max 1GB on both cards. Also GPU ram does not add to more.

Also the more cards the more heat your computer produces. Maybe instead of buying high end equipment
cloud GPU rendering could be an option. I used in my last client job and it was fantastic - extremely fast!

I love my GTX 570 and are happy with it. Faster than anything I ever had anyway!

You’re welcome :slight_smile:

If you look at the benchmark spreadsheet it seems that performance doesn’t really increase significantly with each new generation, so even 2xx series cards can provide good results. You can also see that some results have entries under the “GPU2” column, those are cases where SLI was used. I’m not sure if memory scales when you add a second card though.

If i could spend 500$ i’d prefer to go for a beefy CPU instead of a GPU though. There are CPUs which can outperform even some of the hi-end graphics cards while at the same time on CPU you can use all the memory you want . Not only that, but experimental features are always supported on CPU first.

ah none of the CPUs outperform the GPUs when using CUDA - what are you smoking ???

I recently built a workstation specifically for blender and put in a Titan. I can only recommend it (if you can handle the cost).

I’m just deriving that from the benchmark results. For example you can compare entry number 86 (GTX650TI - 2 minutes, 23 seconds) to entry number 211 with the i7 3930K (1 minute 41 seconds). Obviously, the fastest times are on systems that have both a hi-end CPU and GPU, but if i had to pick only one of those two i’d go for the CPU. Even if the CPU is 10-20 seconds slower it’s still more bang for the buck for the reasons i mentioned above (memory and compatibility).