Hi, i wanted to check kernel improvement in support of strix halo under Debian GNU/Linux, while latest minor versions of 6.16.x improved GTT wanted to check if can be even better. So i tested it on Debian 13 with latest kernel from testing 6.16.12+deb14+1-amd64, and one precompiled performance optimized kernel 6.17.8-x64v3-xanmod1. I ran tests agains Qwen3-Coder-Q8 in full context, but i did benchmark up to 131k. Llama.cpp versions i used for tests: Vulkan build: 5be353ec4 (7109) and ROCm TheROCK precompiled build: 416e7c7 (1). Side notice i managed to compile finally llama.cpp with external libs from AMD for HIP support, so from now one i will use same build for Vulkan and ROCM. Since i wanted also to find sweet spot in energy efficiency so i tried to capture also power usage, and compare it with computing performance. So in the end i tested that model with two backends, and kernels, changing context in few steps, to find out.

In the end seems that latest kernel from testing 6.16.12 works just great! Performance kernel speed is maybe fraction better (max 2%). Besides stock kernel had 4W in idle (in balanced mode), while performance kernel had always minimum 9-10W. And i use fans with 0RPM <= PWM 5% so it's completly silent when idle. And audible under heavy load especially with ROCm. Anyway most optimal power setting for computations is latency-performance and it's not worth to use accelerator-performance in the long run.

Here just notice for strix halo Debian users (and other distros probably too, but current Arch and Fedora have newer kernel), you need to use at least 6.16.x to have better experience with that platform. For Debian GNU/Linux easiest way is to install newer kernel from backports, or move to testing for the latest one. I just noticed that with apt update just now that there is 6.16.12 in stable, so it's great nothing to for Debian users. :) And testing moved to 6.17.8+deb14-amd64 so great, anyway i will have now that kernel, so will test it soon again from debian branch. haha, what an irony, but it took me quite time to write it down. So update: and just tested 6.17.8+deb14-amd64 and idle now is 6W in balance mode now, bit more, than before, but less than the custom kernel.

Performance wise Vulkan is faster in TG, while significantly slower in PP especially with long context. On the other hand ROCm is much faster in PP, and bit slower in TG, but overal improvement in PP is so big that it does not matter for long context (it's around x2.7 faster in 131k CTX window). Vulkan is very fast for shorter chats, but over 32k CTX it's getting much slower. Under load (tested with accelerator-performance profile in tuned) ROCm can draw around 120W (this backend use also more CPU for PP), while Vulkan peak was around 70W.

I found that best values for -ub batch size is 512(it's default) for Vulkan, but 2048 for ROCm (it's faster ~16% than default). After that you have to increase -b logical batch size to 8192 for best performance with ROCm. For Vulkan just leave default logical batch size.

BONUS section, agent test: After tests i wanted to check Qwen3-coder-Q8 model in some tooling so i tried to install kubectl-ai, and connect it to my local llama-server, and perform some tasks on local kubernetes (4 nodes). Model was able based on the natural language promp install Jupyter hub from helm charts, using ~50k tokens for that. And one could run notebooks in some 8-10 minutes. That model works really good on strix halo, worth to check if you didn't yet.

I hope someone will find it valuable, and diagram clear enough. :)