For each of major vendor's latest GPU architectures, is there a clear maximum "triangles/second" bottleneck? If so, what is it architecturally and what is the performance?
GPU Architecture: What are the theoretical limits on triangle throughput in modern GPUs?
Your question is not that easy to answer, as there is no standard definition of what "triangles/second" actually means.
What shading is used? What calculations are done? How are the triangles represented? This and much more has to be taken into account when looking for benchmarks on GPU-performance.
To actually compare the performance of GPUs you would have to look at a range of different measurements. GFlop/s and memory-speed as some basic ones, but also look at full benchmark scores (of which there are I am sure many).
There are websites that compile these measurements and display them in a comparison mode. A quick search brought me to this page which seems to have everything needed (I entered an example search).
This helps to compare the performance of two GPUs but if you want to know if a specific GPU is able to do the specific thing you want it to do, that may be very hard to answer and impossible without much more knowledge about what you want to do.
Calculating the maximum performance is also very hard because I believe the GPU vendors do not share everything they do on their GPUs so it can be hard to estimate the actual performance.
2$\begingroup$ Your point is well taken as a matter of practice: when I used to work at AMD, to test architectural limits, I would do things like clock down the GPU so I could simulate "infinitely" fast memory. However, such architectural limits are all part of a well designed, balanced GPU; understanding gives insight into GPU design, implementation and programming. $\endgroup$– user2500Jan 23, 2016 at 15:58
NV Maxwell: According to the NVIDIA GeForce GTX 680 white paper, each Streaming Multiprocessor (called an SMX in Kepler) in the Kepler architecture could process 1 polygon every other clock. The 680 has 8 SMX units, yielding 4 polygons/clock. The NVIDIA GeForce GTX 980 white paper only mentions that geometry performance of 980 should be double that of the 680; the 980 has twice as many Streaming Multiprocessors (called a SMM in Maxwell), implying each Maxwell SMM has the same polygon throughput per clock as a Kepler SMX. The number of SMMs in a particular GPU isn't always easy to find, but the number of cuda cores is usually publicized; since Maxwell has 128 cuda cores per SMM (there are 192 in a Kepler SMX), we can conclude that Maxwell based GPUs can handle 1 polygon/clock for each 256 cuda cores.