Glad it’s been useful. Just for a bit more context, as a rare AMD RDNA AI/ML guy (see my general AMD GPU doc) I was approached a while back to help give the machines a workout on the AI/ML enablement/profiling front. Bugs and reports were were originally published just as “Strix Halo” (that was released in Feb and with lots of mobile and now desktop chips running the APU), but I was given the go ahead to mention that this was on pre-prod Framework Desktop hardware last month (but also requested to leave all non-LLM stuff to the reviewers, which is fine by me, I already have an actual job 
).
BTW, a while back I started (but never finished) this doc LLM Inference Benchmarking Cheat‑Sheet for Hardware Reviewers to potentially help HW reviewers since very them (including most of the people that focus on AI/LLMs, surprisingly, seem to have much idea about how inference works (eg, the relationship between compute, memory bandwidth, memory size, quants, and model architectures have on inference), or the best ways to get repeatable and reliable results, etc.
In general, testing methodology is still pretty rough (for home hardware, I’m not expecting anyone to report TTFT or ITL, or do concurrency sweeps, but most aren’t even running anything very repeatable or relevant to local AI enthusiasts). That’s where hopefully this repo will hopefully come in handy: GitHub - lhl/strix-halo-testing - the llm-bench provides a simple script for bs=1 (the most common local LLM inference use case) that will let any generate arbitrary pp/tg sweeps across different backends and configs for different models.
I’m actually re-running some tests now atm btw - the latest Vulkan (amdvlk 2025.Q2) and ROCm (TheRock 7.0 nightly, rocWMMA HEAD) have shown some pretty decent perf improvements just in the past month or so. I do wonder if any media reviewers will get this testing right, as there are easily 2-4X differences based on backend, compile/runtime flags, and with kernel/driver configs. (Of course, all my testing is for Linux, I have no idea how things are on the Windows side).