AMD Ryzen AI Max+ 395 Mini PCs

The AMD Ryzen AI Max+ 395 is the chip that pulled workstation-class computing into the mini PC form factor. Code-named Strix Halo and introduced in 2025 as the flagship of AMD’s Ryzen AI Max 300 series, it pairs 16 Zen 5 cores and 32 threads (base clock 3.0 GHz, boost up to 5.1 GHz) with an unusually large integrated GPU and a wide memory bus. AMD builds it on TSMC’s 4nm process, and it lands in mini PCs far more often than the laptop-first framing of its spec sheet would suggest.

What separates this chip from an ordinary laptop APU is the graphics and memory design. The integrated Radeon 8060S carries 40 RDNA 3.5 compute units running at 2900 MHz, and the memory controller is a 256-bit LPDDR5X-8000 interface rather than the narrow single or dual-channel setup on efficiency-class parts. AMD supports up to 128GB of that LPDDR5X as unified memory shared between the CPU and GPU, and its Variable Graphics Memory feature can assign up to 96GB of the pool to the GPU as VRAM (leaving roughly 32GB for the operating system and CPU-side work). That single design decision is why the 395 became the local-LLM favorite: a model that will not fit on a 24GB discrete card can sit entirely in GPU-addressable memory here. An on-die XDNA 2 NPU adds up to 126 TOPS for on-device inference, and the default TDP is 55W with a configurable range of 45W to 120W.

In mini PCs, the 395 shows up in models like the Beelink GTR9 Pro, GMKtec EVO-X2, Bosgame M5, Minisforum MS-S1 Max, HP Z2 Mini G1a, and the Nimo Ryzen AI Max 395. Most ship with soldered memory and dual M.2 slots, competing on cooling, networking, and how much of the 128GB ceiling they offer. This hub covers what the chip does well and where a cheaper part serves you better.

Performance

On raw CPU throughput the 395 sits in a different tier from other mini PC silicon. Its PassMark CPU Mark of roughly 55,005 is about 55 to 60 percent higher than both the AMD Ryzen AI 9 HX 370 (35,497) and the Intel Core Ultra 9 285H (34,478), the next-fastest parts you tend to see in this category. Sixteen full Zen 5 cores give it multi-threaded headroom that reads like a desktop.

The graphics gap is even larger. The Radeon 8060S posts an integrated G3D mark around 18,074, nearly three times the Core Ultra 9 285H’s integrated graphics (6,615) and more than twenty times an Intel N100 (760). In practice that puts the 8060S in the range of an entry-to-midrange discrete GPU, enough for genuine 1440p gaming with no separate card in the system.

For local AI, the story is bandwidth and capacity rather than peak speed. The 256-bit LPDDR5X-8000 interface works out to a 256 GB/s theoretical peak, far above any ordinary laptop chip but still well below a discrete card’s dedicated VRAM. That trade shows up by model size: on smaller models a discrete GPU is faster, but on large models the 395 wins because it can actually hold them. Mixture-of-experts models in the 30B class run at comfortable conversational speeds, and even large dense models can load fully resident in GPU memory, where a 24GB discrete card would fall back to slow CPU offloading or fail to load at all.

Ideal Use Cases

Local AI inference is the headline use. The 96GB VRAM ceiling lets you run large language models, image and video generation, and multi-model pipelines on your own hardware without cloud dependencies. Some owners even link two units over fast networking to pool memory for a model a single box cannot hold.

Beyond AI, the 395 is a capable creative and gaming machine. The Radeon 8060S handles 1440p gaming without a discrete GPU, and the 32-thread CPU paired with 32GB or more of memory chews through 4K video editing, 3D work, compilation, and virtualization, a legitimate single-box replacement for a desktop workstation.

Where it is overkill is ordinary desktop life. Web browsing, office work, media playback, and always-on home-server duties do not touch the GPU, the NPU, or the wide memory bus, and a far cheaper N100 or N150 mini PC will do those jobs quietly and for less power. Buy the 395 for the workloads that actually use its silicon.

What to Look For

Memory is the first decision, and it is permanent. The LPDDR5X is soldered and cannot be upgraded, so pick your capacity (commonly 32GB, 64GB, 96GB, or 128GB) at purchase. If local LLMs are your reason for buying, go as large as the budget allows, since model size is set by this choice.

Cooling and acoustics vary widely between chassis. Because the configurable TDP runs up to 120W, how a given mini PC sustains and cools that power (and how loud it gets) differs by model, so check the specific unit. Networking is the other differentiator: some models ship dual 2.5GbE while others offer 10GbE, which matters if you plan to move large datasets or cluster machines. Finally, confirm the port and storage layout you need, particularly USB4 and M.2 slots. This is a premium tier of mini PC; it earns its place only when your work leans on the GPU, the NPU, or that large memory pool.