A normal desktop PC uses dual-channel RAM at best (2 * 64-bit). But "Strix Halo" is a 256-bit chip/APU, with easily double the RAM speed (easily, because of 8000MT/s RAM, on top of the quad-channel). Why does "Strix Halo" exist? Because AMD specifically advertises "Strix Halo" for AI/LLM usage, basically double the speed of a desktop PC, but also compact. Nothing a normal, dual-channel, 128GB RAM, desktop PC couldn't do, it's just faster (and more compact). 1900$ for only 128 GB RAM is pretty expensive tho. Mainstream AMD AM5 B850 motherboards support up to 256 GB RAM, so they could fit a bigger LLM file, but at that point the speed wouldn't be great. Don't forget, if you have a lot of input text, you may want to add a dedicated NVIDIA GPU for (much) faster prompt processing (pp) and increases token generation (tg) (an AMD GPU may also work, but may be not hassle-free).

AMD, when Strix Halo successor with double the memory bus width (=512-bit) and double the RAM (=256 GB RAM)? 128 GB RAM is very limited and simply not enough; it's not even that fast: 256-bit @ 8000 MT/s = 256 GB/s theoretically and 210 GB/s practically.