Right. Say -- pessimistically -- that I'm on a 5k display (14Mpixels) at 8 bytes per pixel, and double buffered, we can do about 35 copies of that display in GPU RAM. This seems about in line with my experience.

And this is where I don't know what to think because "keep it all in vRAM" is clearly a terrible algorithm, and I have faith in the macOS developers, and yet my data suggests that this is exactly what they're doing. I don't see things swap out when there's memory pressure; it just keeps on allocating until it's full, then spills over to system memory and makes everything slow.

Counterintuitively, the iGPUs have an advantage here. They always run from system memory but they have a faster path to system memory -- they don't need to operate over PCIe -- so (a) you don't have a large perf gap between the 'full' and 'not full' cases, and (b) spillover doesn't hurt. You use a lot more RAM, but you can buy more of that. I can't easily add RAM to my GPU. At best I can double it, and that just means 70 Chrome tabs instead of 35.

I'd love to hear from someone at Apple who's seen the internals of this.