Series 01In progress / 2026
Local LLMs
on Strix Halo.
A product investigation from the hardware up: model fit, real throughput and the small decisions that decide whether a powerful local AI machine is genuinely useful.
Why this machine
Memory changes the product conversation.
Strix Halo combines Zen 5 CPU cores, integrated RDNA 3.5 graphics and up to 128 GB of fast system memory. On the top configuration, AMD allows a large share of that memory to be assigned to graphics workloads.
That does not automatically make it a great LLM product platform. Software support, memory bandwidth, prompt processing, power and thermals still matter. This series is about the gap between can run, good to use and worth deploying.
Platform details are based on AMD’s published Ryzen AI Max+ 395 specifications. Test-machine details will be recorded in each result.
Publication roadmapUpdated as the work moves
Six notes, one reproducible path.
-
01Published
GMKtec EVO-X3: the baseline
A hardware inventory, provable firmware state and repeatable NVMe test from the theoretical link ceiling to a real model-file read.
-
02Planned
ROCm without folklore
A clean install, the working versions, the unsupported corners and a recovery path when the stack moves.
-
03Planned
llama.cpp: Vulkan vs ROCm
Prompt processing, generation speed, memory pressure and the flags that materially change behaviour.
-
04Planned
Finding the useful quant
Where model quality, context length and responsiveness meet for 30B, 70B and larger classes.
-
05Planned
The long-session test
Sustained generation, temperatures, clocks, power draw and whether the first result survives hour two.
-
06Planned
A local stack worth keeping
The final everyday setup: models, runtime, API layer, remote access, backups and honest limitations.
The test contract
Every result will ship with its context.
- 01Exact model and quantisation
- 02Runtime, build and launch flags
- 03Prompt-processing and generation rates
- 04Memory use and context length
- 05Power mode and sustained thermals
- 06Anything that failed on the way