AMD Ryzen AI MAX+ 395 "Strix Halo" — Llama.cpp Backend Performance Comparison

Compare model throughput across backends (pp512 & tg128). Repo: kyuz0/amd-strix-halo-toolboxes

Loading meta…

Search model

Quant

Model params

4B50B96B143B189B235B

4B – 235B

Flash Attention

Winner = every selected backend within the best’s uncertainty range, combining ± errors from both results.

Prompt Processing (pp512) — tokens/second