trying to squeeze every bit of performance out of llama.cpp, clear linux os? anyone has the max tok/s set up on cpu?

[ouvaa]

as titled. 1tok/s extra is worth the switch. also trying to get new rig for the 281gb new mistral. seemed like only amd motherboard with 1tb ram etc can make this work. gpu is too expensive considering the power consumption price factor.

i would prefer to future proof the rig to be able to run 1tb llama models in future.

anyone has any such setups to share? especially with operating systems used.