Beginners: What Card should I buy ?

What is the cost effective way (used card?) that I can buy to play arround and start looking into the following topics.

I would like to use it to learn option and derivate pricing calculation and also to try KI and Deep Learning.
What Cuda Revision would I need at least or which Card would you reccomend.
I would have a budget of 500, but if its absolutly nessessary a bit more.

Thanks for your feedback.

Use CUDA 7.5, which is the latest released version, and has relative few bugs (based on personal experience and judging by bugs reported in this and other online forums). If you want to go cheap on the hardware, any Maxwell-generation GPU will serve you well, and used ones should be very affordable at this point. I haven’t done recent market research so don’t know where the performance/price sweet spot is right now. A handy list of GPUs based on the Maxwell architecture can be found on Wikipedia:

since you goal is to learn CUDA, there are no strict perfromance limits. OTOH, RAM size may be important since it allows to run algorithms with larger requirements.

With $500, you have to choose between consumer GPUs. There are two latest generations - GeForce 9xx and 10xx, and they all should be inside your budget except for GeForce 1080. GeForce 9xx cards are cheaper, but less energy-efficient and has a bit more performance gotchas. So, GeForce 1070 with 8 GB of RAM looks like the best choice. The only alternative may be (most probably used) Titan X (Maxwell) with 12 GB RAM - it’s 9xx series card but with faster+larger memory than 1070.

There are also cheaper cards which are pretty good yet - 1060 (6 GB), and (used or fresh) 980 Ti (6 GB).

Unless I am missing something, any realistic use of a Pascal-based GPU would require the use of CUDA 8.0 RC, which I could not, in good conscience, recommend to a CUDA beginner.

I have to assume important reasons have prevented NVIDIA from shipping the final release of CUDA 8.0. It has been dragged out for months now, considering the historical delays between RC and final versions. If CUDA 8.0 were generally in excellent shape, the final release would have shipped by now.

Why CUDA RC cannot be used for learning purposes? I feel that it has minor problems that novice hardly can encounter, and Release will be made before he will go into deeper fields

Overall, i see not much difference. Real choice, probably, is between used 9xx card and CUDA 7.5, and new 10xx card and CUDA 8 RC

For a beginner, I would always recommend something proven and stable but relatively recent (Maxwell supports all modern features that make a CUDA programmer’s life easier, and is power efficient), rather than a still somewhat experimental platform like Pascal + CUDA 8 RC, where NVIDIA is still working out the hardware and software kinks. Basically, remove any unnecessary complexity and risk to make the learning curve as smooth as possible. For someone on a budget, being able to buy “last years model” cheap(er) should also be attractive.

For seasoned CUDA programmers (I would suspect you fall into that category :-) going with the cutting edge may be OK, but even there I see pain and frustration in the current situation with Pascal + CUDA 8 RC.