What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The most reproducible setup for a Ryzen AI MAX+ 395 mini PC is Ubuntu 24.04 LTS with ROCm 7.2.1 and Ollama 0.20.x. Install the matching AMD graphics stack, pull a quantized model such as Qwen3.5 35B, and verify with ollama ps that the layers are actually on the Radeon 8060S GPU. The platform’s 128GB unified-memory design can hold models far larger than a conventional integrated GPU, but the default 64GB GPU-accessible allocation, context length, operating-system use and CPU/GPU split determine whether a model is fast or merely fits.
This is a versioned AMD example, not a universal guarantee for every mini PC, BIOS revision or later software release. Windows has different runtime choices, and models above the GPU-accessible allocation generally run with mixed CPU and GPU loading.
As an Amazon Associate I earn from qualifying purchases.
What the Ryzen AI MAX+ 395 platform can—and cannot—do
The Ryzen AI MAX+ 395 combines a high-end mobile CPU with Radeon 8060S integrated graphics and a large LPDDR5X unified-memory pool. In the configuration used for AMD’s published inference results, the machine had 128GB of LPDDR5X and 64GB configured as GPU-accessible memory. CPU and GPU draw from the same physical memory, so this is not equivalent to installing a separate 128GB graphics card: the operating system, applications, model weights, key-value cache and runtime all compete for that pool.
That distinction explains most expectations. A 35-billion-parameter mixture-of-experts model can be fully GPU-resident and responsive, while a 122-billion-parameter model may fit in total memory but spill a substantial portion to the CPU. More capacity lets you load larger models; it does not ensure that every layer runs on the GPU or that generation will be fast.
#1 Best Overall
- Spacious Design: Measuring 21.1" wide and 14.1" deep, our lap desk comfortably fits most laptops up to 15.6". Extra room for accessories ensures convenience.
- Enhanced Functionality: Packed with handy features, including a 5x9" precision tracking mouse pad and a built-in phone slot for seamless work or video calls. Plus, enjoy ergonomic support with the integrated cushioned wrist rest.
- Cool Comfort: Enjoy a stable surface with our lap desk's dual bolster cushion, designed for comfort and airflow, keeping your lap cool during extended use.
- Durable Surface: Work with confidence on our lap desk's solid surface, featuring a sleek black carbon color, ensuring optimal air circulation to prevent your laptop from overheating.
- On-the-Go Convenience: With an integrated handle and lightweight design (2.8 lbs), our lap desk is portable for travel or moving around the house, offering flexibility in any space.
AMD describes the benefit this way: “With 128 GB of unified memory, the entire memory pool is accessible for AI workloads — enabling models with 100 billion or more parameters to run on a single system, with no second card and no cloud bill.” That statement describes capacity. Practical speed still depends on quantization, context length, memory allocation, offload and the particular runtime.
A representative mini PC
TechRadar’s 2026 review of the GMKtec EVO-X3 covered a machine with this processor, Radeon 8060S graphics and a reviewed 128GB unified-memory configuration. The review found local AI use possible but said that users needed knowledge and experimentation to unlock the platform. Do not assume every retail configuration has 128GB; check the exact memory option, firmware and cooling design before buying.
Choose an operating system and runtime
| Path | Best fit | Backend and interface | Important qualification |
|---|---|---|---|
| Ubuntu 24.04 LTS + ROCm + Ollama | Repeatable command-line setup and local API use | ROCm, Ollama service and terminal commands | AMD’s documented baseline uses ROCm 7.2.1 and Ollama 0.20.x on a Radeon 8060S system. |
| Linux + ROCm llama.cpp | Users who need direct control of GGUF files, context, batching and offload | ROCm through llama.cpp and llama-server |
Use AMD’s current device and operating-system selector; required libraries and supported GPU/APU targets are version-specific. |
| Windows + LM Studio | Graphical model management and a local OpenAI-compatible endpoint | LM Studio over llama.cpp, with Vulkan or ROCm runtime choices where supported | Update the AMD driver, verify support for the exact machine, and re-check current runtime availability. |
| Windows + prebuilt llama.cpp | People comfortable downloading GGUF files and launching a server manually | AMD’s Windows ROCm package, including gfx115x targets for the Strix Halo family |
The server requires a separate GGUF model and explicit context, GPU-layer and Flash Attention settings. |
For a first installation, Ubuntu with Ollama has the clearest documented smoke test. Windows can be more convenient if you want a GUI, but its supported ROCm components have changed by release. AMD’s Ryzen limitations notes for ROCm 7.2.1 stated that PyTorch was available on Windows while the rest of the ROCm stack was supported on Linux at that time. Treat that as a release-specific snapshot and check the current compatibility notes before committing to a Windows workflow.
Prepare the hardware and memory allocation
Confirm the machine’s real specification
- Processor: AMD Ryzen AI MAX+ 395.
- Graphics: Radeon 8060S integrated GPU.
- Installed memory: the published example uses 128GB LPDDR5X; lower-memory models exist and should be treated separately.
- GPU-accessible memory: 64GB in AMD’s example, configurable on the platform.
- Cooling and power: sustained inference can run the CPU and GPU together, so a stable power adapter and adequate ventilation matter even though no special accessory is required by the software setup.
There is no universal BIOS menu name for the memory control. Use the firmware or vendor utility for your particular mini PC to inspect the AMD Variable Graphics Memory or equivalent setting. Increasing the GPU allocation can make a larger model fit entirely on the GPU, but it leaves less immediately available to the operating system and other programs. Change one setting at a time and record the original value so you can recover if the system becomes unstable.
Rank #2
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Plan memory before downloading a model
A quantized model’s file size is only the starting point. Runtime buffers, GPU workspaces and the key-value cache for the conversation consume additional memory. Longer context windows increase that cache, sometimes enough to turn a model that loaded successfully into one that swaps or fails during generation. Close browsers, containers and other memory-heavy applications when testing the largest models. Set a context length appropriate to the model and task instead of selecting the maximum value automatically.
Install the documented Linux baseline
The following sequence mirrors AMD’s published environment: Ubuntu 24.04 LTS, ROCm 7.2.1 (HSA Runtime 1.18), Radeon 8060S graphics and Ollama 0.20.x. Package names and supported devices can change, so use the current AMD installation instructions for the exact operating system and APU before installing ROCm.
1. Install Ubuntu and the matching AMD stack
- Install or boot Ubuntu 24.04 LTS and apply its normal updates.
- Install the ROCm release that lists your Ryzen AI MAX+ 395/Radeon 8060S device as supported. Follow AMD’s current Linux device and operating-system selector rather than copying a package command from an unrelated GPU guide.
- Install the system libraries required by ROCm and reboot when the driver instructions request it.
- Verify that the runtime can see the Radeon device before involving an LLM. If the device is absent, fix the driver, permissions or supported-device issue first; an Ollama reinstall will not repair a missing GPU runtime.
2. Install Ollama and pull a first model
AMD’s example uses the Ollama installer and a Qwen3.5 model:
curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3.5:35b
ollama run qwen3.5:35b
ollama psThe first command installs the service. The pull command downloads the model, which can take substantial time and storage. The run command starts an interactive session; after it has loaded, open another terminal and run ollama ps. AMD says the process list should report 100% GPU when all layers are offloaded to the Radeon 8060S.
Rank #3
- Note: Not suitable for MacBooks released after 2023 or devices with a protruding front camera; Not applicable to full-screen or notch-style tempered glass screen protectors; Do not use on the rear camera of the phone.
- 💻 Why Do You Need a Webcam Cover Slide? — Safeguard your privacy by covering your webcam with our reliable webcam cover when not in use. Don't let anyone secretly watch you. Stay protected!
- ✅ Thin & Stylish — Enhance your laptop's functionality and aesthetics with our 0.027" ultra-thin webcam covers. Seamlessly close your laptop while adding a touch of sophistication.
- ✅ Fits Most Devices — Compatible with laptops, phones, tablets, desktops! Keep your privacy intact on Ap/ple, Mac/Book, iPh/one, iP/ad, H/P, L/novo, De/ll, Ac/er, As/us, Sa/msung devices.
- ✅ 365 Days Protection — Our upgraded 3.0 adhesive ensures a strong hold that won't damage your equipment. Experience reliable, long-term privacy protection day in and day out.
3. Interpret the process list
- 100% GPU: the model layers are fully offloaded according to Ollama’s process display. This is the desired state for the smaller starting model.
- A GPU/CPU split: the model is larger than the available GPU allocation or other memory pressure prevented complete offload. Generation can still work, but CPU transfers usually reduce throughput.
- CPU only: the ROCm device may not be visible, the runtime may have selected an incompatible backend, or the model did not meet the GPU’s requirements.
Do not judge speed until the process state is understood. A low token rate caused by CPU fallback is a setup problem; a lower rate from a deliberately mixed large model is an expected trade-off.
Pick a model using capacity, latency and quality
| AMD-published example | Quantization and size | Placement | Reported generation rate | What it demonstrates |
|---|---|---|---|---|
| Qwen3.5 9B | Q4_K_M, 6.2GB | 100% GPU | 29.84 tokens/s | A small model with fully local GPU execution. |
| Qwen3.5 35B-A3B | Q4_K_M, 20.5GB | 100% GPU | 42.04 tokens/s | A mixture-of-experts model can be faster than the smaller dense example because only 3B active parameters are used per token. |
| Qwen3.5 122B-A10B | Q4_K_M, 76GB | 61% GPU / 39% CPU | 8.59 tokens/s | A model can fit in unified memory yet exceed the 64GB GPU-accessible allocation and require mixed loading. |
These figures are AMD ROCm Blogs measurements published in 2026 on the disclosed setup. They are not guaranteed minimums. Differences in context length, prompt size, thermal limits, background applications, firmware, quantization build and runtime version can change the result.
For interactive chat, begin with a model that stays fully GPU-resident. Move to a larger model when its quality is worth the latency penalty, and watch the process placement after every model or context change. A mixture-of-experts label such as “35B-A3B” contains both total and active parameter counts; total parameters describe capacity and storage, while active parameters help explain per-token compute.
Use llama.cpp when you need direct controls
llama.cpp works with GGUF model files and exposes controls that are useful when Ollama’s defaults are not enough. On Linux, install ROCm for the selected device, obtain a compatible GGUF file, and build or install a llama.cpp package that includes the ROCm backend. AMD’s installation guidance lists supported Radeon GPUs and Ryzen APUs plus required host libraries; use its live selector for the current command sequence.
Rank #4
- Anti-Slip Surface - Transform your laptop into a mobile workstation with the AboveTEK portable laptop lap desk. The anti-slip surface provides a strong grip for laptops up to 15.6 inches(Diagonal), while the double rubber strip on the bottom ensures a stable display or typing experience on your lap, couch, or bed.
- Retractable Mouse Pad - Retractable laptop mouse pad extends on both directions for the left/right handed with elevation along the edges for stopping mouse from falling off. The size of laptop tray is 14" X 9.7" and the size of mouse pad is 7.4" X 6.1".
- Effective Heat Shield - The effective heat shield made of sturdy and thick material protects your laptop from overheating. Prioritizes your comfort and safety, an ideal lap pad or board for working anywhere.
- EASY to Carry and Store - With an ergonomic and simplistic design, the lap desk is portable to store in a backpack. Only 15" in size, 2.2 lb of weight and with slim 0.6 inch thickness, it is ready to be easily carried around.
- Widely Applicable - The smooth platform accommodates laptops and tablets up to 15.6 inches(Diagonal), making it a versatile accessory and one of the best gifts for mom, dad, students and professionals. Perfect for use as a laptop bed tray or tablet holder anywhere at home, library, or park.
On Windows, AMD provides prebuilt ROCm packages for supported architecture targets, including gfx115x for the Strix Halo architecture family identified in those packages. You still need to download a GGUF model separately. A typical server launch conceptually includes:
llama-server -m path/to/model.gguf -c 8192 -ngl 999 -faThe exact executable name and flags depend on the package version. In this example, -c 8192 chooses an 8,192-token context, -ngl 999 requests aggressive GPU-layer offload, and -fa enables Flash Attention when the build supports it. Confirm the current flag names in the installed version. If the model is too large, reduce the requested GPU layers or context rather than assuming the command has failed.
llama.cpp’s server mode is useful for applications that speak an OpenAI-compatible local API, but endpoint settings, authentication and network exposure are your responsibility. Keep a local server bound to the local machine unless you deliberately configure access controls.
Free tools Windows power users keep installed
One-click scans. No signup required.
Set up a Windows graphical workflow
LM Studio provides a graphical interface over llama.cpp. AMD describes it as able to download and deploy models, select Vulkan or ROCm runtimes for AMD where available, and expose an OpenAI-compatible local endpoint for other applications.
Best Value
- Spacious Design: Measuring 21.1" wide and 12" deep, our lap desk comfortably fits most laptops up to 15.6". Extra room for accessories ensures convenience.
- Enhanced Functionality: Packed with handy features, including a 5x9" precision tracking mouse pad and a built-in phone slot for seamless work or video calls. Plus, enjoy laptop support with the integrated device ledge.
- Cool Comfort: Enjoy a stable surface with our lap desk's dual bolster cushion, designed for comfort and airflow, keeping your lap cool during extended use.
- Durable Surface: Work with confidence on our lap desk's solid surface, featuring a blush pink color, ensuring optimal air circulation to prevent your laptop from overheating.
- On-the-Go Convenience: With an integrated handle and lightweight design (2.14 lbs), our lap desk is portable for travel or moving around the house, offering flexibility in any space.
- Install current AMD graphics drivers for the mini PC and reboot.
- Install LM Studio and allow it to identify the available backend. If both Vulkan and ROCm are offered, test the backend that is documented for your exact driver and device rather than assuming one is universally faster.
- Download a quantized model in a format supported by the selected llama.cpp runtime, normally GGUF.
- Choose a conservative context length and enable GPU offload in the model settings.
- Start a local chat, then inspect the runtime’s load information to confirm whether layers are on the GPU.
- If you change AMD Variable Graphics Memory, reboot before retesting; the allocation change is not reliably applied to an already running session.
Windows support is particularly version-sensitive. AMD’s ROCm 7.2.1 limitation notes did not describe the same breadth of Windows support as Linux, so verify the current release notes for the runtime you intend to use.
Troubleshoot failed or slow inference
The GPU does not appear
- Confirm that the Ryzen AI MAX+ 395 and Radeon 8060S are listed as supported by the installed ROCm release.
- Check that the operating-system driver and user-space ROCm packages come from compatible releases.
- Reboot after driver installation or a firmware memory-allocation change.
- Run the runtime’s device query before launching Ollama or llama.cpp. If it cannot see the device, troubleshoot the platform installation rather than the model.
Ollama reports CPU use
- Run
ollama pswhile the model is active and record the placement. - Try a smaller quantized model that is known to fit within the GPU-accessible allocation.
- Close memory-intensive applications and lower the context length.
- Check that Ollama is using the intended ROCm-capable runtime and that no incompatible package has replaced it.
- Do not increase GPU allocation blindly: taking memory from the operating system can create a different failure.
The model loads but generation is very slow
First distinguish prompt processing from token generation and check whether the model is mixed CPU/GPU. The 122B example’s 8.59 tokens/s is a useful illustration of the cost of a 61% GPU / 39% CPU split, not a universal benchmark. Large contexts, thermal throttling and background memory pressure can also lower throughput. Reduce context, use a smaller quantization or choose a model with fewer active parameters when interactive latency matters.
The process is killed or the system becomes unresponsive
This usually indicates that total memory, not just GPU-accessible memory, was exhausted. Stop the model, close other applications, restore a less aggressive GPU-memory allocation if necessary, and restart. Then load a smaller model or context. A model file that appears to fit on disk can still require substantial runtime and cache memory.
Recommended Free Tools
A practical first-day workflow
- Record the mini PC’s exact memory capacity, firmware version and current GPU-accessible allocation.
- Choose Ubuntu 24.04 LTS for the documented baseline, or confirm the current Windows runtime matrix before selecting Windows.
- Install the matching ROCm/driver stack and verify the Radeon device independently.
- Install Ollama, pull Qwen3.5 35B, and run the
ollama psoffload check. - Measure one short, repeatable prompt while noting model quantization, context length, placement and temperature.
- Try a smaller 9B model for lower memory use, then a larger model only after you understand the machine’s headroom.
- For models that exceed the GPU allocation, decide explicitly whether the quality gain justifies mixed CPU/GPU latency.
- Use llama.cpp or LM Studio when you need GGUF-level controls, a graphical interface or a local OpenAI-compatible endpoint.
What this mini PC is best suited for
A fully GPU-offloaded 9B or 35B-class quantized model is the practical sweet spot for conversational use, coding assistance and document work where response time matters. The 128GB configuration is valuable when you need a larger model, a longer context or several models stored locally, but those benefits are capacity benefits. A 76GB model that leaves only part of its layers on the GPU may deliver better answers at a pace unsuitable for rapid back-and-forth.
The platform is therefore best viewed as a flexible single-box inference system: more capable than ordinary integrated graphics, quieter and simpler than a multi-GPU workstation, but dependent on careful memory and software configuration. Treat AMD’s published numbers as a baseline for reproducing the environment, not as a promise for an unverified mini PC or a later release.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
