In general, you should choose a workstation-class (Nvidia Quadro or ATI FirePro) or a gamer-class (Nvidia GeForce or ATI Radeon) graphics card. 
If you want to use Deep Learning with Dragonfly, you can only choose Nvidia cards with CUDA support. Please see this page for more information: 
While Dragonfly performance will improve with more video RAM (VRAM) and more GPU processing capability (GPU cores), there is no certain answer for which card is sufficient for you, as it depends on the size of your data files and what you are going to do with those files. For many users of micro-CT and 3D-EM, a card with 4G VRAM is about the minimum requirement to get the work started, and 8G VRAM or more is highly recommended. 
Below are the highest-performance graphics cards to consider, ranked from highest to lower performance (last updated February 2019), where the last four-digit number in each description, for example "4608", indicates the number of CUDA parallel-processing cores. 
For the same amount of VRAM and GPU capability, the GeForce cards are usually lower priced than the Quadro cards. For example, the Nvidia GeForce RTX 2070 8GB  card costs about 500$ (as of May 2020), making it a good choice for low budget.
  • Nvidia Quadro RTX 8000 48GB 4608 (Turing)
  • Nvidia Quadro GV100 32GB 5120 (Volta)
  • Nvidia Quadro RTX 6000 24GB 4608 (Turing)
  • Nvidia Titan RTX 24GB 4608 (Volta)
  • Nvidia Quadro P6000 24GB 3840 (Pascal)
  • Nvidia GeForce RTX 2080Ti 11GB 4352 (Turing)
  • Nvidia Quadro RTX 5000 16GB 3072 (Turing)
  • Nvidia Quadro GP100 16GB 3584 (Pascal)
  • Nvidia Quadro P5000 16GB 2560 (Pascal)
  • Nvidia GeForce RTX 2080 8GB 2944 (Turing)
  • Nvidia Quadro RTX 4000 8GB 2304 (Turing)
  • Nvidia GeForce RTX 2070 8GB 2304 (Turing)
  • Nvidia Quadro P4000 8GB 1792 (Pascal)
When building a system yourself, you must confirm that your system is compatible with your preferred graphics card. The constraints are the interface of the motherboard (e.g. PCIe 3.0) and the power capacity of the power-supply unit (e.g. 800 W).
For more information: