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

If you can’t beat ‘em, join ‘em

AMD Ryzen AI Halo
(Image credit: © Tom's Hardware)

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.

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.

AMD Ryzen AI Halo

(Image credit: Tom's Hardware)

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.

TOPICS
Jeffrey Kampman
Senior Analyst, Graphics

As the Senior Analyst, Graphics at Tom's Hardware, Jeff Kampman covers everything that has to do with graphics cards, gaming performance, and more. From integrated graphics processors to discrete graphics cards to the hyperscale installations powering our AI future, if it's got a GPU in it, Jeff is on it. 

  • Neilbob
    Me when trying to understand the meaning and purpose of such devices (and A.I. in general).

    LKCi0gDF_d8
    AMD just blowing bubbles, right along with N****a. I expect Intel will be next up with a meaningless overpriced plastic box.
    Reply
  • Pierce2623
    Software compatibility is a negative compared to an ARM based machine?? Sure guys…. Also it seems only benchmarking some AI models and nothing else is purposely catering to Nvidia. AI hadn’t even started booming when we first heard about Strix Halo. So it clearly wasn’t designed solely for AI regardless of how they’ve advertised during the AI boom. If the AI boom hadn’t happened Strix Halo would be getting sold with 32GB for $1000. Knowing all that context, it seems weird to only test a few AI models.
    Reply
  • Bigshrimp
    What are these devices for the price? I know this is a rhetorical question, but it's still funny that this was released. Costly for what it is...
    Reply
  • suryasans
    AI writing by Google Gemini is much smarter and fairer than you. https://www.google.com/search?q=amd+ryzen+ai+halo+ai+developer+pc+vs+nvidia+gb10+power+consumption&sca_esv=4b04a474141781e0&rlz=1C1CHBF_enID1038ID1038&sxsrf=APpeQnsquHCA-jw-qqP98xNQhxFXMchnpg%3A1783360176978&ei=sOpLaqiwO_SVseMP0ty0qAw&biw=1600&bih=731&oq=amd+ryzen+ai+halo+ai+developer+pc+vs+nvidia+gb10+power+co&gs_lp=Egxnd3Mtd2l6LXNlcnAiOWFtZCByeXplbiBhaSBoYWxvIGFpIGRldmVsb3BlciBwYyB2cyBudmlkaWEgZ2IxMCBwb3dlciBjbyoCCAAyBRAhGKABSOhOUKIJWIk8cAF4AJABAJgBhwGgAasFqgEDNy4yuAEDyAEA-AEBmAIKoALIBcICCBAAGO8FGLADwgILEAAYgAQYogQYsAPCAgUQABjvBcICCBAAGIAEGKIEmAMAiAYBkAYFkgcDOC4yoAedE7IHAzcuMrgHxQXCBwM1LjXIBwyACAE&sclient=gws-wiz-serp
    Reply
  • maviz
    This is a completely useless piece of trash on a piece of shit ecosystem.
    I had the Ryzen with 32 GB and i will tell you the drivers are *USELESS*.
    Even above that, the compatibility matrix is laughable...i had to sell both my CPU and my 6700 XT...and now i have a 5060 that is leaps and bounds faster and "just works"....and another 5070 TI which was not comparable to begin with. This box is hopelessly slow for any real AI work, priced like shit, performing like shit, i dont even know how they dare present this as any sort of evolution. You spend a few hundred more and get something 3 yrs ahead in the ecossystem and also MANY times faster...its laughable at best.
    Reply
  • JamesJones44
    Pierce2623 said:
    Software compatibility is a negative compared to an ARM based machine?? Sure guys…. Also it seems only benchmarking some AI models and nothing else is purposely catering to Nvidia. AI hadn’t even started booming when we first heard about Strix Halo. So it clearly wasn’t designed solely for AI regardless of how they’ve advertised during the AI boom. If the AI boom hadn’t happened Strix Halo would be getting sold with 32GB for $1000. Knowing all that context, it seems weird to only test a few AI models.
    AI training and inference is all done on Linux and largely with ARM based servers believe it or not (they cost less and the CPU isn't all that relevant for AI workloads). The fact that x86 lags in this department isn't a surprise.
    Reply
  • dva852
    Pierce2623 said:
    Software compatibility is a negative compared to an ARM based machine??
    AI is a CUDA world, at least for now. ROCm support is still spotty. x86/ARM is irrelevant.

    AMD is leaning on open-source angle as it tries to breach CUDA moat. It also places higher focus on Windows platform, gunning for Windows devs as they move to AI. Nvidia is doing same with RTX Spark. Windows itself is moving to ARM. Lots of moving pieces.
    Pierce2623 said:
    Also it seems only benchmarking some AI models and nothing else is purposely catering to Nvidia.
    You don't buy a $4K box to play games on an iGPU. Strix Halo was already considered "overpriced" at $2K--which admittedly was before-DC era. But repurposed for AI, it has gained new relevance, and higher valuation. IMO, not $4K high, but "higher."

    At $4K vs DGX Spark's $4.7K, Strix AI is underwhelming. DGX has more compute, faster interconnect, better ecosystem. But the price is a bit irrelevant. It's a reference device, with official software (Win & Lin) stack, and most importantly, official support. It's meant to establish AMD's AI ecosystem. Think of it as placeholder for future iterations.

    Even though Spark has faster compute, both it and Strix Halo are saddled with same memory bandwidth bottleneck. It's why the models used in AMD's benchmark (as well as THW ones) are either mixture-of-expert (read: low active parameter count), or small dense models. Even a midsized dense model would've brought both to a crawl. That's the real takeaway, that LPDDR5X is a makeshift solution to the AI bandwidth problem, and will need better solutions going forward.

    Gorgon Halo also on LPDDR5X. Medusa Halo gets LPDDR6/X, in either late '27 or '28.

    I looked on YT and "Ryzen AI Halo reviews" are popping up, so apparently today is "Halo AI's" official coming out party. Underlining the point is AMD's YT blurb, with Jack Huynh's mug trying reeeeaally hard to crack a smile. More happy thoughts, Jack.

    For a more informative Ryzen Halo AI review, as well as providing a wider perspective,

    Gz62bniDkpgView: https://www.youtube.com/watch?v=Gz62bniDkpg
    Pierce2623 said:
    AI hadn’t even started booming when we first heard about Strix Halo. So it clearly wasn’t designed solely for AI regardless of how they’ve advertised during the AI boom.
    After a forgettable foray in some high-end laptops, Strix Halo popped up in Shenzhen mini-PCs, and was immediatedly co-opted for AI use. The review below was 9 months ago, and gives a good perpective on Ryzen AI's WIP state at that point in time. Note also the $1.8K price.

    prIUKAbHlj8View: https://www.youtube.com/watch?v=prIUKAbHlj8
    Reply
  • salgado18
    Neilbob said:
    Me when trying to understand the meaning and purpose of such devices (and A.I. in general).

    LKCi0gDF_d8
    AMD just blowing bubbles, right along with N****a. I expect Intel will be next up with a meaningless overpriced plastic box.
    For coding, for example: you set up a coding agent platform, like Claude Code, Codex or OpenCode. Then, you ask it to implement something or fix a bug. The platform will orchestrate AI agents and local tools to understand, plan, execute and test the task. That's the use case.

    Why such a box? Cloud-hosted AI models cost a lot of money if charged per token, and monthly subs have limits and are subject to price increases or vanish entirely. To run locally, even most gaming PCs struggle with such a heavy task that is running AI, let alone many of them in parallel. And setting up a central server with good hardware is crazy expensive and a big maintenance task.

    Enter these boxes: you pu one on your desk, work on it or connected to it, and the agent platform uses its hardware to work, instead of cloud models. So your work is entirely local and contained to one PC.

    I believe this is the future of PCs, and we're just seeing the first prototypes and proof-of-concepts. It's terrible value for gaming, regular working, media editing etc. But for local AI usage, it's a very tempting format.
    Reply
  • salgado18
    dva852 said:
    "For coding, for example: you set up a coding agent platform... and then it's immediately you spin up a containerized runtime, pipe your LLM through a RAG vector store, scaffold the MCP tool schema, YAML the CI/CD hooks, kubectl apply to the Kubernetes cluster--

    FTFY o_O
    What? I tested Claude Code and OpenCode, and it's just install -> set api key -> start prompting. Even with Ollama it's quick and easy.
    Reply
  • JamesJones44
    The crazy part is this is starting to look like somewhat of a bargain. A year ago you could have gotten an 128GB M4 Max Mac Studio with double the memory bandwidth of the max+ 395 for $3700, now to get that level of performance you are in the 5.5+k range with GB 10 or Apple Silicon. Performance for local model operations is probably good enough for low to medium intensity tasks.
    Reply