GuideLocal AI
Ryzen AI Max+ 395 mini PC or NVIDIA DGX Spark? What each is really good at
Two 128GB AI mini computers compared on speed, software and what you will actually use them for.
Published 24 Oct 2026
You want to run a very large model at home, one that no single graphics card can hold, and two small boxes promise 128GB of shared memory to do it. They look alike on paper and behave quite differently.
Picture the problem. The model you want is too big for any consumer graphics card, so you look at the new breed of small computers that share 128GB of memory between processor and graphics. Two names come up again and again: AMD's Ryzen AI Max+ 395, found in mini PCs such as the GMKtec EVO-X2 with 128GB, and NVIDIA's DGX Spark. Tests that compared them show the answer depends on what you plan to do with the box.
Which is faster for AI?
For generating text from a model that is already loaded, they land close together, and in The Register's test the AMD machine was slightly ahead on one software path. The gap opens with long prompts. The DGX Spark began answering two to three times sooner on a short prompt, and the difference grows as the prompt gets longer. One review measured prompt processing on a 120 billion parameter model at about 340 tokens a second for the AMD chip against about 1,700 for the Spark, roughly five times slower. Across fine-tuning and image generation the Spark was also ahead, by about two to three times overall.
How big a model can the AMD mini PC run?
About 96 to 110GB of the 128GB is usable for AI, which is why it can load models far beyond any single graphics card. Speed depends on the type of model. A mixture-of-experts model such as gpt-oss 120B runs at about 34 tokens a second, close to the Spark's roughly 38. A dense 70 billion parameter model is slower, about 5 to 15 tokens a second depending on how it is compressed, because memory bandwidth, about 215 GB/s in practice, limits it.
What does the DGX Spark do better?
It runs NVIDIA's CUDA, so most AI software works without fuss, and it is built as a dedicated AI appliance for prototyping agents, fine-tuning models and generating images or video. It also costs noticeably more.
What does the mini PC do better?
It is a normal computer. It runs Windows or Linux, handles everyday work and games as well as AI, and costs less. The catch is software. It uses AMD's ROCm, Linux is the faster path, and tools such as vLLM and Flash Attention needed building from source or special versions in the test.
Is the built-in NPU useful?
Not yet for big models. On the NPU a 7 billion parameter model ran at only 4 to 5 tokens a second against an expected figure near 40, so treat it as a bonus.
Which should you choose?
- One machine for work, games and chatting with large models, especially mixture-of-experts ones: the Ryzen AI Max+ 395 mini PC.
- A dedicated AI box for fine-tuning, long prompts, image or video generation, or CUDA-only software: the DGX Spark.
- Mostly models under 32B: a single graphics card is faster and cheaper than either.
Sources
AMD Strix Halo vs Nvidia DGX Spark: Which AI workstation comes out on top? by The Register · Dec 2025 - AMD Ryzen AI Max+ 395 for Local LLMs (2026): Is 128GB Unified Memory Worth It? by HashTechWave
Written by Startbook from the sources above and our own research; each belongs to its publisher. Prices on linked products are live and may differ from those in the sources.