According to many, NVIDIA RTX 2080 Ti is the best GPU for deep learning, when price-performance ratio is accounted for. It costs around 1K (customized editions cost more) and you can parallelize up to 4 cores with SLI, depending on your motherboard, of course. However, multiple GPU sets come with their own disadvantages at some regards. There's always NVIDIA Tesla V100, that costs a lot more (9K I guess), but offers an overall faster performance compared to a single RTX 2080 Ti, without all the hassle associated with multi-GPU configurations. That is, of course, if you're looking for the most convenient way. This article contains useful informtion:

https://timdettmers.com/2019/04/03/which-gpu-for-deep-learning/

I'm not a professional, so I would personally go with a single V100. However, if you're adept enough to work with multiple GPUs, you can have four 2080s with half of the price.

https://developer.nvidia.com/taxonomy/term/696

NVIDIA RTX 2080 Titan is the best GPU for image processing and deep learning.

Currently, only Nvidia GPUs are only are compatible with deep learning frameworks. I am using Nvidia GTX 1080. Nvidia GTX 1080 Ti is also another option.

You can also go for Nvidia RTX 2080 Ti as suggested in other answers.

If you need the GPU temporarily, consider using cloud services like :

https://lambdalabs.com/service/gpu-cloud

which is much easier to set up and maintain.

You can try using Google Colab which provides GPU and it's totally free.

https://colab.research.google.com/notebooks/intro.ipynb