128 GB, $3,999, and 200B Parameters: AMD's Ryzen AI Halo Reviewed

Reviews of the AMD Ryzen AI Halo landed this week, and the picture that emerges is of a capable local-AI workstation at $3,999 — roughly half the price of NVIDIA's competing DGX Spark — with 128 GB of unified memory accessible as both system RAM and VRAM. For anyone running large language models locally, those numbers matter.

The Core Idea: One Pool for Everything

The Ryzen AI Halo is a compact desktop box built around AMD's Ryzen AI Max+ 395 (Strix Halo) chip. The key design choice is unified memory: all 128 GB of LPDDR5X-8000 is shared between CPU, GPU, and NPU, delivering roughly 256 GB/s of bandwidth with no memory-copy overhead between components. Up to 96 GB of that pool can be assigned as VRAM.

In practical terms, you can load and run models with up to roughly 200 billion parameters on this machine without aggressively quantizing the weights — something not feasible on typical consumer GPU setups capped at 24 GB of VRAM.

What the Reviews Actually Found

Phoronix called it an excellent and powerful mini PC with fully open-source software support — no proprietary driver stack required. Running models purely on the NPU, Flopper's review recorded a peak of just 35 W total chip power, which is remarkable efficiency for the class of task.

The ceiling is memory bandwidth. At 256 GB/s, the bus is faster than consumer alternatives but slower than NVIDIA's DGX Spark — reviewers cite this as the main reason the Halo trails NVIDIA on inference throughput for the largest models. HotHardware notes that ROCm on Windows still needs polish; on Linux the experience is more complete. The machine excels at LLMs but struggles with video generation, which is more bottlenecked by GPU throughput than memory capacity.

The Value Calculation

NVIDIA's DGX Spark starts at $4,699 for the 128 GB configuration (some vendors list it at $7,000). At $3,999, the Ryzen AI Halo saves you at least $700 while running entirely open-source software — which matters considerably to AI researchers and the hobbyist community.

If LLMs are your primary workload and Linux is your OS, the Halo is a strong contender. If you need NVIDIA's CUDA ecosystem or Windows-first tooling, the price gap may not be enough to compensate for the ROCm rough edges.

Why It Matters

A year ago, running a 70B-parameter model locally required serious compromises or five-figure spending. The Ryzen AI Halo at $3,999 is not cheap, but it's the price point where individual researchers and small organizations start treating local inference as a real alternative to API costs — and that has implications for privacy, latency, and who can afford to experiment at the frontier.