Ok, thanks for that. I was under the impression that PRT+ was more granular than PRT by streaming parts of mips
Terminology around some of these things is in flux (and consequently a bit of a mess).
Best I can recall, PRT came into active use around 10 years ago (in game contexts anyway), while streaming on mip-map granularity has been around for over 20 (again, in games), so those concept were never meant to be interchangeable. The whole point of "partially resident" is that any given mip-level isn't completely loaded, so it always referred to finer granularity (eg. VRam pages, or smaller).
Best way to contextualize SFS is 'hardware features used
by PRT' - eg. Rage implemented sampler-feedback through a simple software rasterizer, allowing it to ship across entire stack of DX9 GPUs of the era. But obviously - at the cost of some CPU time. And while this may seem like a problematic concession, it's worth remembering that cutting edge engine-tech of 2007-ish era used software rasterization to also generate occlusion coverage, force propagation and more - so you'd get multiple uses from going down this route.
As for the multipliers - that's a really poot way of explaining the benefit(though it's obvious why they chose it). The key point is that any given frame in a 4k render accesses less (way less actually) than 512MB of unique data, so if you can
optimally fetch from SSD(ie. what SFS should allow to get close to) - that's all the Ram you'd ever need for 'asset storage'. Makes those 16GB suddenly look much bigger.
So what component of the GPU shared on RDNA 2 PC cards, nvidia cards, and XSX/S APUs is responsible for hardware VRS tier 2?
It isn't necessarily the same component, VRS doesn't dictate implementation details, and there are multiple viable approaches to the problem. But all of them require certain extensions to existing hardware, think of it less like 'new component' and more like how RDNA2 extends texture-sampling to accelerate intersection math for RT for example..
news.xbox.com
