3 ms·
The point of shared memory is that you can skip a whole memcpy. You load from drive into shared RAM and use it immediately. Preferably it's stored on disk in th
by vinyl7 3y ago
The point of shared memory is that you can skip a whole memcpy. You load from drive into shared RAM and use it immediately. Preferably it's stored on disk in the same format that the GPU can use so you don't have to parse.
Compare that to a PC with a discrete GPU, you have to load the texture/mesh into RAM and (possibly memcpy to a mapped buffer that the GPU can see) and then issue a copy to move it over PCIe to dedicate GPU vram.
- kllrnohj 3y agoThat only marginally improves asset streaming performance. Both CPU RAM and a pcie x16 slot have much more bandwidth than any nvme drive will (which is only x4 pcie, after all, even ignoring the limitations of the NAND flash & controller itself) so that extra memcopy isn't significant. And has absolutely no impact whatsoever on general VRAM usage as was claimed anyway. Also you don't have to load it to system RAM first. DirectStorage exists (kinda)