I'm not sure if Kitty is using GPU specifically to minimize latency, but it would be a good reason.
That being said I think it is valuable to have a reasonably fast terminal because it's not uncommon to have applications slow down dramatically if they output a lot of information on stdout/stderr while the terminal struggles to render everything.
Input was just so .... immediate? It’s really hard to describe, but it was surprisingly obviou.
I can definitely tell you that while I can't see the actual letter appearing on the screen - the letter on the calculator would feel as if it appeared right at the bottom position of the button, while on my PC the key would have travelled some distance off the ground before I see the letter.
It's not uncomfortable, but I could easily tell the two apart in a blind test would something like that come up.
It's depressing that he's only ever tried for throughput, when that's irrelevant next to latency. And it does have abysmal latency as per your link.
I've been using iterm for years; when they introduced their metal rendering backend the impact was very subtle but noticable. Anyway, given the amount of time I spend using cli stuff, it kind of matters.
Smooth scrolling is inherently bounded by the vsync rate of the display. As long as you can render the terminal's pixel buffer with new characters and send it to front for the next flip, all within some portion of the time you have between frames, it's as smooth as it can get. Missing a frame will cause stutter, and make it non-smooth. Given the fact that terminal doesn't necessarily need to render if there are no changes, using GPU seems like an overkill for regular use.
Text rendering is basically just blitting cached glyph bitmaps into a buffer, and CPUs have been more than overly fast doing that for eons. And CPU rendering has none of the compatibility problems/quirks that GPUs have. I can fire up a Gnome Terminal onto an unaccelerated Vesa X11 display if my graphics card doesn't work or I haven't configured it yet.
For better responsivity I suppose a terminal that behaves like mosh would be a good approach. It would maintain the tty state in memory and only render the latest view, dropping the rendering of any intermediate diffs in the buffer. I would guess most terminal emulators do work like that these days but I haven't checked.