AMD Ryzen AI Halo review: AMD builds a DGX Spark of its own

Nvidia’s DGX Spark and its GB10 SoC have set the template for what a purpose-built local AI developer sandbox should be. The combination of a standardized hardware platform with robust first-party software support and thorough documentation lets those curious about local AI get up and running faster than buying a bare-metal box and building everything up from scratch, especially in the rapidly evolving AI space.

AMD’s Ryzen AI Max+ 395, aka Strix Halo, SoC, is the best x86 spoiler for GB10 so far. It has the same 128GB of unified memory, a powerful 16C/32T Zen 5 CPU, and a Radeon 8060S integrated GPU with 2560 RDNA 3.5 stream processors. It also has an AMD XDNA 2 NPU for those who want to experiment with that accelerator in addition to the general-purpose Radeon GPU. And it can run Windows and Windows apps natively, whereas GB10 boxes are Linux-only for now.

AMD’s partners have been building around this hardware for about a year and a half, and it’s a well-known quantity at this point. But once you have that hardware in hand, setting it up for AI workloads involves digging through scattered GitHub pages, Reddit threads, and AMD official documentation to get all the software pieces lined up right for the best performance and compatibility.

AMD is trying to change all that today with the launch of the Ryzen AI Halo, a first-party, turn-key Strix Halo mini-PC that puts local AI first. This system can be had with Windows or Linux, and at least in the Linux form we’re testing today, it comes preloaded with the full AMD ROCm software stack and an assortment of applications you need to immediately start generating tokens with your preferred model.

And on the support side, AMD has taken a page directly out of Nvidia’s book and cooked up an entire set of its own playbooks that cover various local AI applications and usage scenarios with the AI Halo (and Strix Halo systems more generally) to serve as a springboard for local AI explorers.

The grand tour

The AI Halo comes wrapped in a plastic shell with a subtly color-shifting finish. It’s got a large light bar ringing its front and sides that indicates system status. White means it’s awake, while a pulsing blue indicates that it’s asleep, assuming you allow it to suspend at all. Red indicates a fault. If you find the LED strip distracting, you can just turn it off using the preinstalled AI Developer Center app.

The AI Halo has air intakes on its top and sides, and AMD cautions that you shouldn’t block any of these intakes. If you’re running by the book, that means this system is less flexible than it could be for space-constrained or multi-node home lab setups, where turning the unit on its side would allow for valuable space savings.

Enterprising community members will likely design and share 3D-printed spacers and risers to get around these limitations, but for a device that is presumably meant to be used in home labs and production environments, the lack of flexibility in orientation is a small but annoying oversight.

Around back, the AI Halo has the same trio of USB Type-C ports you’ll find on Nvidia GB10 boxes, plus one more for power input with the included 240W brick. The port closest to the power plug runs at “USB 3.2” speeds, while ports 3 and 4 are higher-speed USB 4. These ports are all DisplayPort Alt Mode compatible, or you can use the HDMI 2.1 port for display output if you prefer.

For wired networking, the AI Halo offers a 10 Gigabit Ethernet port. That’s certainly fast, and AMD has written a clustering playbook for multiple AI Halos using that interface, but it’s in a whole other league compared to the 200Gbps ConnectX-7 NIC on the DGX Spark and its ilk.

We didn’t want to strip our AI Halo all the way down to its guts, but each of the four rubber feet on the bottom of the system is secured with a pair of tiny magnets, and they conceal the four screws you presumably need to remove to get further inside.

Here’s a quick look at this system’s specs:

Amid the ongoing RAMpocalypse and NANDpocalyse, no Ryzen AI Max+ 395 system with 128GB of RAM and a large SSD is cheap, assuming you can find a 128GB config in stock anywhere.

Even against that backdrop, the $3999 price tag for the AI Halo that we’re testing today is a pricey proposition. That sticker puts it at the low end of Nvidia GB10 systems like the Asus Ascent GX10 (albeit in its 1TB config).

Our past testing of Strix Halo versus GB10 for local AI workloads has decisively put Nvidia’s platform on top, so this is a potentially awkward place for the AI Halo to land. Let’s dig in and find out if anything has changed.

• MORE: Best Graphics Cards
• MORE: GPU Benchmarks and Hierarchy
• MORE: All Graphics Content

I was hoping to expand both the set of inference engines and AI models tested for this review, but I quickly ran into issues. The Ryzen AI Halo ships with a version of vLLM pre-installed, but it wasn’t compatible with Qwen 3.6-35B-A3B when I tried to launch it, so I fell back to the old reliable llama.cpp.

We tested three models using llama.cpp: Qwen 3.6-35B-A3B, a relatively lightweight mixture-of-experts model that’s been extremely popular of late, Google’s Gemma 4 12B, another recent and relatively lightweight but dense model that activates all of its parameters per token, and gpt-oss-120B, a larger mixture-of-experts model that’s been available for quite some time now. We used Unsloth’s GGUF versions of these models in their Q4_K_M quantizations.

This time around, we’re using llama-benchy as our benchmark harness. llama-benchy lets us get llama-bench-like performance results out of any model runner that can present us with an OpenAI-compatible endpoint, not just llama.cpp. That’s quite handy for comparing performance across inference engines.

First up, we’ll look at latency and throughput performance with Qwen 3.6-35B-A3B:

And with the dense Gemma 4 12B:

And finally, with gpt-oss-120B:

At least with llama.cpp, the AI Halo’s relative performance in single-user LLM serving versus GB10 isn’t any different than what we saw several months ago when we took the Corsair AI Workstation 300 through its paces. While its tokens-per-second throughput is fine, albeit still slower than the Dell Pro Max GB10, its time-to-first-token latency falls far behind GB10 as context length grows.

In a real-world scenario where llama.cpp can use (and is configured to use) its prompt caching features, these differences might not be as pronounced, but it’s still important to note this worst-case behavior, especially for long-running coding workflows where context lengths can quickly grow.

Overall, the Ryzen AI Halo doesn’t work any magic for Strix Halo inference performance compared to other implementations of this platform we’ve tested. While its tokens-per-second throughput is acceptable, its time-to-first-token latency can quickly rise to non-interactive levels with long contexts.

At the extremes of our testing, waiting two to four minutes for a model to start responding might be disruptive to an interactive workload like a coding assistant, even if the rate at which tokens flow is tolerable once they do start rolling.

I also ran the same check-in on ComfyUI image generation performance using the same basic Flux.2 Klein test we ran back in February, as well, and the same gulf that we saw in time-to-completion for ComfyUI work remains now. The GB10 GPU’s larger shader complement chews through image generation far quicker than the Radeon 8086S.

Finally, we checked in on CPU performance with Geekbench 6. The AI Halo’s 16-core, 32-thread Zen 5 CPU holds an edge on the GB10’s 20-core Arm CPU complex in both single-threaded and multi-threaded performance, so tasks like code compilation could potentially be faster on the AI Halo. But that victory shouldn’t cause us to lose sight of the GB10’s all-around better AI performance.

Thermal performance and noise levels

We didn’t want to tear down our Ryzen AI Halo to reveal its cooling system, but as we discussed at the beginning of this review, the system has vents on its front, top, and sides to allow for plenty of airflow, and you can see a decently sized copper heatsink through its rear vents.

Logging system temperatures on Linux is more difficult than it is on Windows, but we didn’t see CPU or GPU temperatures higher than the mid-50 °C range when running the AI Halo through our typical workloads. All that suggests that this box is more than up to the task of cooling the chip inside.

As for noise, the Ryzen AI Halo runs its twin blower fans audibly even at idle, so it’s always adding some amount of noise to a room. And under a ComfyUI generative workload, it gets significantly louder than the Dell Pro Max GB10 box we’re using to represent that platform.

The noise signature of those fans is also a bit less refined than those in the GB10 boxes I’ve used. They have a notable high-pitched whine that’s difficult to acclimate to when the system is placed on a desktop, whereas the GB10 systems I’ve used all just sound like moving air and are easy to ignore.

• MORE: Best Graphics Cards
• MORE: GPU Benchmarks and Hierarchy
• MORE: All Graphics Content

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

AMD’s Ryzen AI Halo is the company’s attempt to put forth a complete, turn-key, first-party hardware and software package for AI developers, backed by direct software support and an extensive library of documentation for running local AI tasks on the Strix Halo platform. This effort unsurprisingly resembles the hardware, software, and documentation ecosystem that Nvidia has built around the DGX Spark.

Despite those lofty goals, the included software configuration on the AI Halo doesn’t necessarily put the best foot forward for this platform yet. On the upside, the included Lemonade local AI sandbox, which is the focus of a lot of AMD optimization for local AI workflows, is easy enough to get started with, has a polished user interface, and performs well. Of all the beginner-friendly experiences on this box, Lemonade is perhaps the most polished. And the AMD Sync application gives you handy remote access to this system’s processing power from anywhere you can tunnel into it.

We also appreciate the handy AI Developer Center hub for its one-stop system management options, but we wish it allowed us to dig into tasks like managing local models. AMD’s configuration choices for preinstalled apps like ComfyUI also made it harder for us to work with them as shipped versus simply downloading and installing them ourselves.

So even if you are truly starting from zero local AI knowledge and need a completely on-rails experience, you’re likely to find yourself chafing at AMD’s configuration choices sooner or later. And while the provided playbooks are broadly useful, they’re sometimes missing key information, like folder paths, that makes deeper exploration of the concepts within difficult.

In any event, you don’t need an AI Halo to access AMD’s playbooks, so if you’re curious about the quality of this documentation, you can review it independently before you buy – or use it with any other compatible Strix Halo box.

All this is important because the AI Halo is a significant investment. At a list price of $3999, the AI Halo is about 16% cheaper than the DGX Spark as of this writing. AMD touts this lower price as a cost-per-token advantage, and that might be appealing for the well-heeled hobbyist or enthusiast who just wants to play in a local AI sandbox.

But if you’re a developer for whom time and tokens are money, you can get into an Asus Ascent GX10 GB10 system with a 1TB SSD and 128GB of RAM for the same $3999 as the AI Halo. And even the $4700-ish price tag of a DGX Spark with its 4TB SSD will pay for itself in fairly short order simply because it keeps you waiting less.

Especially if you’re trying to learn the ropes of local AI work, the quality of Nvidia’s accompanying documentation and the breadth of its application support is still better than what AMD has shown so far for the AI Halo, so you’re more likely to have a smooth ride.

All told, our verdict for the AI Halo is mixed. This is certainly the most turn-key Strix Halo box available for local AI work, and if you’re all-in on the AMD AI ecosystem, want a direct line from AMD for software support and documentation, highly value the ability to boot both Windows and Linux, potentially need to play with AMD’s XDNA 2 NPU, and don’t mind lower performance than a DGX Spark in exchange for all of those options, then maybe the AI Halo is for you.

But there’s no two ways about it: this box is still generally slower and less agile as an AI development platform than a GB10 system, and until AMD is ready to ship a next-gen SoC in the architectural shape of Strix Halo with RDNA 4 graphics or some other future GPU IP, that value proposition looks like it’s going to be very difficult to shift.

• MORE: Best Graphics Cards
• MORE: GPU Benchmarks and Hierarchy
• MORE: All Graphics Content