Why you can trust Tom's Hardware Our expert reviewers spend hours testing and comparing products and services so you can choose the best for you. Find out more about how we test.
Given its performance deficit versus Nvidia's GB10 platform, the next question for the AI Halo is whether AMD's included software and library of documentation adds substantial value versus other Strix Halo boxes that ship with nothing more than an OS.
I started my exploration of AMD's preinstalled software and playbooks with Lemonade, a heavily AMD-backed and AMD-optimized sort of software Swiss army knife for AI inference that makes it easy to play with a range of models, inference engines, and modalities, all in one unified interface.
While I didn't dig too deep into its capabilities, I was able to quickly combine Qwen 3.6-35B-A3B with the llama.cpp backed and start chatting with it, all with performance similar to what we saw in our directed testing with llama.cpp. This is the kind of smooth, straightforward experience that you want from a product like this. But Lemonade is available for a wide range of AMD systems, not just AI Halo, so you don’t have to buy one of these boxes if you want to try it out.
Other included apps have a few more wrinkles. When you first boot the Linux AI Halo we tested, the system launches a centralized management interface called the Ryzen AI Developer Center (AIDC for short) that aims to put key system information, settings, and software updates in one easy-to-access spot.
The AIDC app exposes some handy settings that Strix Halo users will want at their fingertips. Most notably, it lets you adjust how much of the AI Halo's 128GB memory pool is split between the CPU and GPU graphically instead of through the command line.
While we didn't have time to test their impact on performance, you can also select from three power profiles to balance noise, power consumption, and performance. And if the built-in LED light bar on this box is annoying you for any reason, you can turn it off through this interface.
Outside of those functions, the AI Developer Center has some good ideas that don't feel fully baked. For example, model management is one of the most common tasks when I'm setting up or mucking around on a local AI box. The AIDC app will show you what models you have downloaded across the system for use with AMD's pre-installed apps. But you can't open or otherwise reveal their containing folders so that you can add to or manage the data within them directly.
That's a pain, because if you do want to extend the usefulness of preinstalled apps like ComfyUI by adding new models in support of different workflows, that task is harder than it should be. As far as I can tell, the preinstalled ComfyUI lives in a Podman container (presumably for ease of updating), so its traditional directory structure (which would normally end up under /home//) is obscured.
ComfyUI can be configured to look for additional model directories beyond its defaults using a separate .yaml file. But with AMD’s configuration, those directories live in /var/cache/, which on the AI Halo’s Debian image is owned by root, not the user, so you can't just drop new models into those alternate directories at will without some chmod work that I didn't want to mess with for fear of breaking something.
And this directory location isn't documented in the related ComfyUI playbook that AMD provides (but is discussed in the overall user guide for the system).
AMD says the preinstalled software and models on the AI Halo are only meant to support its playbooks, and that users are free to go about installing and configuring the same apps to their own taste. But to me, the next logical step for learning after running through those playbooks is extending the functionality of those workflows, and if the process for doing so is neither straightforward nor transferable to the natively installed version of the application, then is the knowledge being conferred even useful?
For another example of some minor documentation hassles, the preinstalled version of vLLM on this system has a launch script that provides a lightweight wrapper that presents status information and checks the health of the vLLM instance. AMD encourages running different models with vLLM by changing the model string in this script, but the path to it is again not documented anywhere in the accompanying playbook. So I had to go digging. I eventually found it in /usr/bin/ by dumb luck, but this is the kind of very basic information that documentation exists to chronicle.
AMD also provides a direct competitor to the useful Nvidia Sync remote access app called AMD Sync. Sync allows you to connect to the AI Halo using SSH, and you can get versions of it for Windows and Linux alike (but not for macOS, so far, an option that Nvidia supports.)
The basic idea of AMD Sync is that you can run workloads that need the resources of the AI Halo on that system while working remotely on your preferred device. For just a couple examples, you can get a terminal session, run a remote instance of VS Code, or work in JupyterLab, all powered by the AI Halo. Check out AMD’s playbook for more examples of what Sync can do.
Overall, the pre-installed software on the AI Halo will get you started if you have no prior local AI experience and are working from AMD's playbooks to learn the basics, but I feel like the configuration decisions that AMD made to include those apps could be more flexible and better documented.
From the playbooks I tried, I also think AMD could stand to have a few more rounds of QA to ensure that these docs fully cover everything beginners need to know in order to get the most out of the platform.
- MORE: Best Graphics Cards
- MORE: GPU Benchmarks and Hierarchy
- MORE: All Graphics Content
