AMD Ryzen AI Halo Review: $4,000 Local LLM Mini PC Faces Memory Bandwidth Limits

What the box is​

TechPowerUp describes the Halo as a machine that handles local AI workloads well. It pairs a Ryzen AI Max+ 395 APU with 128 GB of unified memory in a compact case. The chip is codenamed "Strix Halo." It has a 16-core, 32-thread Zen 5 CPU, an RDNA 3.5 integrated GPU with 2,560 shaders, and an XDNA 2 NPU. The memory is 128 GB of LPDDR5X-8000.

StorageReview reports 256 GB/s of memory bandwidth and a single 120 W USB-C power input. That bandwidth figure matters later, because TechPowerUp lists it as a limit.

AMD also claims the memory pool is enough for models of up to 200 billion parameters. StorageReview relays AMD's statement that the pool can hold models up to that size in device memory. That is a capacity claim from AMD. It says nothing about how fast any particular model will run.

TechPowerUp's value verdict​

TechPowerUp lists the 128 GB Halo with the Ryzen AI Max+ 395 at US$4,000. Its listed strengths are:

• It is much cheaper than the NVIDIA DGX Spark, which the review puts at $7,000 now.
• It runs large models that even a 32 GB GPU can't load.
• It does local AI that is private, with no subscription or cloud cost.
• It ships with AI software preinstalled.
• It supports both Windows and Linux.
• It has 10 Gbit Ethernet and a status LED with fault indicators.

Its listed drawbacks are:

• $4,000 is a lot of money for local AI.
• Cloud-based models can give better answers in many cases.
• It lacks FP8/FP4 support.
• Time-to-first-token is slower than on the DGX Spark.
• The 256 GB/s memory bandwidth limits token rate.
• The fan never stops at idle, and it is a bit loud under load.
• It is limited to PCIe Gen 4, so models load slower from the SSD.

The SSD limit shows up in the review's conclusion, but I could not see the SSD benchmark page behind it.

TechPowerUp frames the Halo as more than a developer kit. It calls the machine potentially a personal AI appliance for power users running big models locally. It also says AMD's AI Playbooks make getting started simple.

How other reviewers see it​

The DGX Spark price is where reviewers disagree most. TechPowerUp uses $7,000. StorageReview puts NVIDIA's own DGX Spark with the larger drive at closer to $4,700, and ties the increase to the LPDDR5X and NAND supply crunch. At the lower figure, the Halo's price advantage mostly disappears. Tom's Hardware reaches that conclusion in its own review, saying the Halo's AI performance trails GB10 systems and that $3,999 is not far below a faster alternative.

Reviewers do agree on the performance trade-off. StorageReview found the Halo beat the DGX Spark on Linux in CPU work, with 7-Zip compression 11% faster and the LLVM compile 14% quicker. It trailed by 2x to 4x in most vLLM serving scenarios at higher concurrency. That fits TechPowerUp's complaint about time-to-first-token and bandwidth.

One roundup also notes a spec discrepancy. TechPowerUp measured the unit at 1,016 grams and the roundup says AMD's spec sheet is wrong on this point. AMD lists the weight as under 1.2 kg, so the two figures may not actually conflict. The roundup's claim is attributed to that source and I could not check it against TechPowerUp's page.

Who should care​

Windows support is a real advantage over Nvidia's offering. Tom's Hardware notes the x86 platform can run Windows and Windows apps natively, while GB10 boxes are Linux-only for now. For developers who need Windows tooling next to local models, that matters.

For most buyers, $4,000 is a hard sell. One independent review calls it a strong official developer box and a weak bargain for an ordinary enthusiast. It adds that the premium pays for support, validation and deployment convenience.

My own read, based on general industry knowledge, is that the Halo suits people who want a quiet-ish, private, plug-in local model server. It is a poor fit for anyone who just wants a fast desktop or a gaming PC. Before buying, check whether the models you plan to run fit in memory. Then check whether 256 GB/s of bandwidth gives you a token rate you can live with.

Bottom line​

TechPowerUp's verdict is that the Halo is a capable local-AI machine that is cheaper than a DGX Spark. Its high price, its bandwidth ceiling and the lack of FP8/FP4 support keep it from being a bargain. The CPU, GPU and SSD test numbers are on the review's performance page, which I could not read. Those figures are worth checking before you buy.