Nintendo Maniac 64 Wrote:That makes me wonder how efb2ram would perform on Kaveri, but I'm pretty sure hUMA (at least as it is in Kaveri) still needs software to expose such functionality, so it wouldn't "automatically" have improved performance...
#dolphin-emu Wrote:(2:04:42 AM) MaJoR: degasus HdkR, do you think the shared CPU-GPU memory pool system that AMD is making will be enough to make up for their crappy processors and make them competitive for dolphin?
(2:08:21 AM) degasus: MaJoR1: nope
(2:08:36 AM) degasus: all gpus support to read directly from main memory
(2:08:51 AM) MaJoR: so why is it slow?
(2:09:37 AM) MaJoR: I assume the separation, which the whole merged system is supposed to address
(2:09:41 AM) degasus: dolphin just generates a huge workload for both, cpu + gpu
(2:09:59 AM) MaJoR: kk
(2:10:09 AM) degasus: we usually don't wait on the gpu but for efb2ram + real xfb
(2:10:38 AM) degasus: so yes, it will speed up efb2ram a bit
(2:11:16 AM) degasus: but we'd still have to wait for all of the rendering, we just don't have to readback the texture twice
(2:12:00 AM) MaJoR: kk
(2:12:27 AM) MaJoR: so EFB to Ram would just run closer to EFB to Texture
(2:12:41 AM) MaJoR: which is already slow on AMD
(2:13:21 AM) JMC47: EFB2Ram runs like shit for me
(2:14:36 AM) degasus: MaJoR1: it will be a small speedup for efb2ram, but as we still have to wait for all of the rendering, it will still be much slower than efb2tex
(2:14:43 AM) MaJoR: kk
(2:15:07 AM) MaJoR: I can't wait for the dashed hopes and dreams of AMD fanboys
Intel Xeon w7-3465X OC | Asus Pro WS W790-E Sage SE | NVIDIA GeForce RTX 4090 FE | 8x16GiB G-Skill Zeta R5 DDR5-6000 | Windows 11 23H2 | (details)
MacBook Pro 14in | M1 Max (32 GPU Cores) | 64GB LPDDR5 6400 | macOS 13