hogeheer499-commits/strix-halo-guide

The current headline data is measured primarily on one Beelink GTR9 Pro. To make the guide more useful, we need comparable results from other Strix Halo / Ryzen AI MAX systems.\n\nUseful systems:\n\n- Framework Desktop 13\n- GMKtec EVO-X2\n- HP ZBook Ultra G1a\n- Beelink GTR9 Pro variants\n- Any Ryzen AI MAX+ 395 / Radeon 8060S system with 64GB or 128GB unified memory\n\nPlease include:\n\n- System model, RAM size, BIOS UMA setting, IOMMU setting\n- OS, kernel, Mesa/RADV, ROCm versions\n- Backend, model, quant, context length, generated token count\n- Exact command used\n- CSV/raw output where possible\n\nBenchmark template:\nhttps://github.com/hogeheer499-commits/strix-halo-guide/issues/new?template=benchmark-report.md\n\nDiscussion thread for early results:\nhttps://github.com//discussions/2

The guide currently does not claim same-machine Windows versus Linux performance. This would be useful for Strix Halo owners deciding whether they need Ubuntu or can stay on Windows.\n\nWanted test shape:\n\n- Same physical machine, same BIOS UMA setting, same power profile where possible\n- Windows Vulkan path: Ollama / LM Studio / llama.cpp if available\n- Linux path: Ollama Vulkan/RADV and/or llama.cpp Vulkan/RADV\n- Same model file or equivalent quant\n- Same prompt length, generated tokens, context length, and repeat count\n\nPlease include raw output and exact commands. If the model files differ, note that clearly so the result is not overclaimed.

The guide has vLLM smoke-test notes, but no headline vLLM throughput claim for a comparable 35B-class workload yet.\n\nUseful targets:\n\n- Qwen3.6 / Qwen3.6-coder class models where available\n- AWQ/BF16/other vLLM-compatible formats\n- Single-stream and small-concurrency OpenAI-compatible serving\n- Tool-calling or long-context notes if supported\n\nPlease include:\n\n- Container/image or build details\n- ROCm/TheRock/runtime versions\n- Model format, quant, context length\n- Exact server command and client benchmark command\n- Raw logs, throughput, TTFT, errors, and startup time\n\nExternal claims are useful leads, but this issue is for locally reproducible results that can be linked from the guide.

The guide intentionally does not publish tokens-per-watt yet because power telemetry is not validated. This issue tracks a reliable methodology.\n\nNeeded:\n\n- Wall power or validated board power measurement\n- Idle baseline and loaded run measurements\n- Same benchmark command used for tokens/sec\n- Notes on power profile, thermals, clocks, and background load\n- Repeat runs or enough samples to avoid one-off numbers\n\nPotential outputs:\n\n- tokens/sec\n- watts above idle\n- tokens/watt\n- total request energy for long-context tests\n\nPlease include the measurement tool/device and raw readings. Until this is validated, README power claims should stay conservative.