FRANKENSTEIN: Strata, B650M, Qwen 3.8 IQXS3, 32K/205K Context, 32GB RAM, PCIE 3.0x4?, MI25's?, 70+Tok/s 700+ Prefill? Yes. It can do that, with a little help.

A hardware tinkerer reported running the Qwen 3.8 IQXS3 model at 32K/205K context on a 32GB RAM system with a B650M motherboard and PCIe 3.0 x4, achieving over 70 tokens per second generation and over 700 tokens per second prefill, and published the configuration and benchmark notes in the GitHub repository onewireout/mi25-llm-inference. The same repository documents testing of AMD Instinct MI25 cards for llama.cpp inference, where the cards fell well short of their theoretical performance, and the author is building a headless inference node using 4x MI25 GPUs.