I created a sphere and subdivided the sphere until it was about 393.5k polygons. Applied the subdivision then exported it into DAO format. I then created a test scene in godot 3.1.1 and stuck the sphere in the test seen, with an fps counter giving results to the console.
Test equipment was a T5500 machine with two 6 core x5660 cpus. They run about 3.2ghz each. For the gpu, running an XFX RX 580 with 4GB ram.
First, I found out that godot was capping the frames per second at 60 fps. I do not know the performance of this test beyond 60 fps.
Each sphere, 393.5k polygons.
I stuck 9 spheres in the scene. The result was about 30 fps. 3.541 Million Polygons.
I stuck 5 spheres in the scene. The result was about 60 fps. 1.967 Million Polygons.
With 6 spheres, I had a result of about 53 fps.
According to steam, the report of the most reported gpu (according to power), is GTX 750 TI - an GTX 1070. This makes up 46.12% of the gpu reported to steam. The GTX 1060 was the most common used GPU with 15% of users.
The Rx 580 is about 55% stronger than a 1060. Not as strong was the 1080, which about 1% of steam users use.
The rx 580 is about 20% stronger than an rx 570. The rx 570 is stronger than a 1060. The 1060 much much stronger than a rx 560. With the rx 560 falling a little bit above the GTX 750 TI
According to my analysis, a polygon count of about 1.65million would give about 60fps on 46% of the gpu reported by steam. I think it would be safe to use about 2million polygons as a maximum target per game scene. This is for high end video cards.
To hit the whole market, not counting the really old stuff. About 800k-1.1 million polygons.
About 50% of steam users have cards slower than the rx 460.
Other machines can do better with the same rx 580 than my test system.
AAA game studios use 11million, on things like the ps4.