Okay, so the reason a framerate limit just under the V-Synced refresh causes a variable jump of up to 1 frame input latency, is because the cap is now introducing frames in uneven/offset intervals of the refresh rate, instead of even intervals without a cap.
As for there being more than 3 frames of V-Sync latency, in the G-Sync + V-Sync on case, we're probably not looking at more than 3 frames, unless of course, the user is either stupid enough to have pre-rendered frames set too high in the control panel, and/or enables in-game triple-buffering on top of the aforementioned combo (which triggers when he hits the G-Sync ceiling), or the case that the game has a high default pre-rendered frames number, and it engages triple-buffering behavior at default with V-Sync enabled.
I guess the question is (yes, I'm starting to sound like a broken record, sorry), is if G-Sync + V-Sync on is possibly mimicking the variable input latency that is introduced in the situation you described, but instead of happening only a fraction of a frame below the refresh rate, it's possible the variable latency is being introduced in our much discussed 125-142 range somehow because of the 1ms polling rate.
I did find this 2013 article from the advent of G-Sync:
http://www.anandtech.com/show/7582/nvidia-gsync-review
Whether or not the module's functions have progressed or changed since then (they probably have), I'm assuming the basics remain the same. I find this portion interesting regarding the V-Blank:
G-Sync works by manipulating the display’s VBLANK (vertical blanking interval). VBLANK is the period of time between the display rasterizing the last line of the current frame and drawing the first line of the next frame. It’s called an interval because during this period of time no screen updates happen, the display remains static displaying the current frame before drawing the next one. VBLANK is a remnant of the CRT days where it was necessary to give the CRTs time to begin scanning at the top of the display once again. The interval remains today in LCD flat panels, although it’s technically unnecessary. The G-Sync module inside the display modifies VBLANK to cause the display to hold the present frame until the GPU is ready to deliver a new one.
With a G-Sync enabled display, when the monitor is done drawing the current frame it waits until the GPU has another one ready for display before starting the next draw process. The delay is controlled purely by playing with the VBLANK interval.
Finally, regarding the module's on-board memory:
The G-Sync board itself features an FPGA and 768MB of DDR3 memory. NVIDIA claims the on-board DRAM isn’t much greater than what you’d typically find on a scaler inside a display. The added DRAM is partially necessary to allow for more bandwidth to memory (additional physical DRAM devices). NVIDIA uses the memory for a number of things, one of which is to store the previous frame so that it can be compared to the incoming frame for overdrive calculations.
Now, this could be pure ignorance coming up here, so bear with me.
From this thread, we can assume that G-Sync + V-Sync off is exposing the supposed 1ms polling rate's inability to complete the frame when the current fps isn't at least 1ms frametime under the max refresh, causing the tearing seen at the bottom of the screen. We also know that G-Sync begins scanning to the display from the very top to the bottom of the screen.
My question to you, is it possible with G-Sync + V-Sync on to use the module's on-board memory to hold frames for the fraction of time the 1ms polling rate needs to catch up in that very small, specific range of, say, 125-142, and delay the render from the top of the screen, to a little lower/higher offset (late or even early?), with the remaining section (where we see tearing instead with G-Sync + V-Sync off) being rendered/synced at a slightly slower or different rate?
I realize my description is severely under-cooked, and may be entirely flawed, but I had to put it out there. Because even if I'm a bit of an uninformed moron (
), it might trigger a bigger truth that those more knowledgeable than I in these matters could gather from it.
Or, someone could perform rigorous latency tests within that 125-142 "G-Sync + V-Sync on" range, and destroy all of this "fun-filled" speculation once and for all.