New to llama.cpp on Strix Halo? This page gives the recommendation in one place; the linked pages carry the full setup steps and the numbers behind it.
The decision table
| If you are… | Use | Why (measured on this machine) |
|---|---|---|
| Running dense models day to day (e.g. Qwen3.8-27B) | ROCm + MTP | ~106% faster generation with MTP than the ROCm baseline; +74% for Vulkan’s MTP — ROCm wins on the top end too |
| Running MoE models at plain decode (e.g. Qwen3.6-35B-A3B) | Vulkan | Vulkan’s baseline generation was ~14% faster than ROCm’s before any speculation |
| Running MoE models with MTP enabled | Either — effectively a tie | Both backends converge to ~60 tok/s with MTP (ROCm 60.9 vs Vulkan 60.4) |
| Installing is fighting you (older Mesa, ROCm packaging issues) | Vulkan / RADV from the Kisak PPA | Fewer moving parts; verified working on this exact hardware |
| Reproducing the numbers on this site | Both builds, same llama.cpp commit | The benchmark runs both backends from one checkout (c060ca974) so only the backend differs |
All figures come from ROCm vs Vulkan: llama.cpp on AMD Strix Halo — a published page with the raw JSON data linked for inspection.
Start here, in order
- Install ROCm 7.14 on Strix Halo —
userspace-only install that keeps the kernel
amdgpudriver and display working. Includes the compiler checks that prove the toolchain targets the right GPU. - Vulkan / RADV compute setup — the fallback backend, from the Kisak Mesa PPA.
- Run llama.cpp on Strix Halo — build both backends, select the GPU explicitly, and read the practical notes: memory behavior, the MTP flag set that was used, and common failures.
- RAM, GTT and GPU memory on Strix Halo — what the 128 GiB unified memory actually is, what the BIOS “VRAM” number means, and how to watch GTT usage while a model is loaded.
Scope and caveats
- These recommendations reflect one machine (Ryzen AI Max+ 395 with 128 GiB
LPDDR5), two Q8_0 models, and one llama.cpp commit (
c060ca974, build 10603) tested in August 2026. Upstream moves fast — after a major llama.cpp, ROCm, or Mesa upgrade, re-test before trusting old numbers. - The MTP win is specific to models with a matching draft head (Qwen3.x here). Other model families may not have one; the plain-decode ranking still stands as the general guide.
- Power profile was Balanced throughout testing. A different profile changes absolute throughput, and possibly the ranking at the top end.
Sources and upstream documentation
- llama.cpp repository — build options for both backends referenced above.
- SPEED-Bench — the methodology behind the MTP numbers on the benchmark page.