With Viewer getting GPU parallelism as seen in the new GPU video on the Gallery page, plus all the expansions like user-specifiable GPU-ized convolution matrix filters, you might find yourself thinking about a new GPU.
I've researched various GPU options, trying to figure out what a good "bang for the buck" solution would be. GPU prices change all the time, so the "bang for the buck" sweet point also changes. A current list of "bang for the buck" inflection points might be:
Card, Cores, Newegg price ($)
GT 710, 192, 40
GT 1030, 384, 85
GTX 1050 Ti, 768, 170
GTX 1060 6GB, 1280, 250
RTX 2070 8GB, 2304, 500
RTX 2080 8GB, 2944, 750
I recommend the GTX 1060 with 6GB (not the 3GB ones). That gives you 1280 CUDA cores for $250 and plenty of on-board memory for GPGPU parallelism. The GTX 1050 Ti with 768 cores at about $170 is also a great "sweet spot" in terms of maximum bang for the buck.
The 710 and 1030 are in there because GIS goes better with three monitors. For that you need two cards. For example, I'm writing this on a rig with a GTX 1060 and a GT 710, a total of 1472 CUDA cores. The 710 is just a cheap way to plug in two more monitors. Manifold will use all those cores optimally regardless of what generation they are on if they are on separate cards.
It also helps to buy the latest generation you can, so you have a longer time before NVIDIA stops supporting the chip with CUDA drivers. Any of the 10x0 or later series are fine.
The really interesting thing we see with the latest builds is how well manycore CPU parallelism compares to GPU parallelism. Unless you are doing lots of really exceptionally math-heavy work, you don't need to feel compelled to go beyond 1000 cores. The main argument for spending $250 on a GTX 1060 6GB, with 1280 cores, is that it is so relatively inexpensive for what it is that it seems a shame not to buy it just in case that extra 6 GB and those extra cores might be useful sometimes.
But I bet in actual use you'd rarely notice any difference from what a GTX 1050 Ti, with 768 cores at a significantly lower price, would accomplish. Heck, even a GT 1030 at well under $100 provides tremendous speedup, and most people won't care if something drops in time required from over a minute to either four seconds or eight seconds.
As always, I recommend doing GIS with three monitors. Plug in a cheap GT 710 or a GT 1030 to drive extra monitors. Manifold will use the extra 192 or 384 cores as well.
If you have the budget, of course, there's nothing wrong with pushing the envelope with a higher end card, like the RTX 2080. Why not go even faster with nearly 3000 cores if you can afford it? I'm just saying that the less expensive cards are also really fast as well.