Minisforum MS-S1 Max Review: The Era of Limited Mini PCs is Officially Over

For years, I've separated desktop PCs into two categories: compact mini PCs for office work and servers on one side, imposing towers for gaming and heavy workloads on the other.

This divide seemed mandatory. A mini PC doesn't run games properly, a desktop tower doesn't fit in a backpack. The Minisforum MS-S1 Max disrupts this logic with its AMD Ryzen AI Max+ 395 processor and 128GB of unified memory.

After a month of testing between gaming sessions, AI model processing, and benchmarks, I'm seriously asking myself: what if compromise is no longer necessary?

This €2,599 machine clearly doesn't target the general public. It addresses AI developers who want to run large language models locally, creators juggling a dozen applications simultaneously, and also gamers who refuse to sacrifice their desk space.

AMD's Strix Halo architecture promises a revolution: making 16 Zen 5 cores and a Radeon 8060S GPU with 40 compute units coexist in a single chip, all with access to 128GB of shared RAM.

Understanding what makes the Ryzen AI Max+ 395 special requires digging into the architecture a bit. AMD borrowed a page from Apple's Silicon playbook. On a classic computer, the CPU, GPU, and RAM live separately on the motherboard. Data must transit through the PCIe bus between these components, creating bottlenecks.

The Ryzen AI Max+ 395 integrates everything into a single package - it's a SoC (system-on-chip): the 16 Zen 5 cores, the Radeon 8060S GPU with its 40 compute units, and especially four LPDDR5X memory controllers instead of the usual two. This configuration doubles the available bandwidth to about 250 GB/s.

The 128GB of LPDDR5X-8000 dynamically shares between CPU and GPU according to needs. You can allocate up to 96GB exclusively to the GPU via BIOS, a VRAM quantity usually only found on professional cards costing €5,000 or €8,000.

The MS-S1 Max is quite attractive. The aluminum chassis exudes solidity. Minisforum didn't try to minimize volume at all costs this time. The machine measures roughly double the height of a typical mini PC while maintaining a similar footprint.

So we could rather speak of a "compact PC" than "mini PC" since the comparison doesn't hold with mini PCs. This size increase isn't an aesthetic whim: it responds to the thermal constraints of a processor capable of climbing up to 120W TDP.

The front face uses a large ventilation grille accompanied by two USB-C USB4 ports, a USB-A 3.2 Gen 2 port, a combo audio jack, and the power button. The system can be positioned vertically or horizontally thanks to anti-slip pads present on two faces.

I particularly appreciate the build quality. The angles are sharp, the ventilation grilles well cut, no play in the assembly. Minisforum has clearly progressed since its first models.

The only annoyance throughout testing: the sticker on top of the case uses incredibly strong adhesive. Three weeks after removing it, my fingers still stick when I touch this area.

The rear connectivity gathers the essentials. Two USB4 v2 in USB-C run at 80 Gbps, a first on a mini PC of this category. These connectors also handle DisplayPort alternate mode and can provide 15W to charge peripherals.

You can connect an eGPU or dock here, like the Thunderbolt 5 Razer Chroma. Here, you can reach 4 GB/s with an external M.2 SSD. Four additional USB-A ports complete the arsenal: two in USB 3.2 Gen 2 and two in USB 2.0. A single HDMI 2.1 output handles the main display.

The two 10 GbE Ethernet ports rely on Realtek R8125 controllers. Under Windows 11, everything works immediately from first boot. Under Linux, the experience varies greatly depending on the distribution. Ubuntu 22.04 and later versions recognize the ports after installing the r8125-dkms package. However, Debian 13 and Proxmox gave me trouble. I had to plug in a USB to Ethernet adapter to continue my tests under these systems.

Disassembly is accomplished by removing two rear screws. The panel then slides effortlessly, revealing the beast's interior. A massive heatsink occupies two-thirds of the available volume. This aluminum structure integrates six copper heat pipes and two 70mm fans. The fins extend along the entire length of the case to maximize exchange surface.

The Lighton 320W power supply sits at the bottom of the chassis. This brand, which disappeared from radars since the CD burner era, reappears with a very decent quality internal block. The presence of an integrated power supply greatly simplifies installation: a simple power cord suffices, with no bulky external transformer.

A dedicated 40mm fan cools the power supply independently of the main circuit. Two M.2 slots accept SSDs in 2280 format. The first, occupied by the supplied 2TB disk, benefits from four PCIe Gen 4 lanes for theoretical throughput up to 7,000 MB/s in reading. The second slot is limited to a single PCIe Gen 4 lane, meaning performance should plateau around 1,750 MB/s.

The PCIe Gen 4 slot in x16 physical format is an interesting addition. Beware however: this connector only has four electrical lanes. This configuration is suitable for adding a 25 GbE network card or a U.2 adapter for professional storage. However, Minisforum explicitly indicates that adding dedicated graphics cards is not supported. You can nevertheless use an external graphics card (eGPU) via one of the USB4 ports.

Let's start with the CPU. Cinebench R24 tests place the Ryzen AI Max+ 395 at the level of high-end desktop PC processors. The multi-core score reaches 1795 points in Performance mode, better than Intel Core i9-13900K and Ryzen 9 9900X which nevertheless consume over 200W.

These 16 Zen 5 cores with their optimized architecture deliver impressive raw power. The single-core score is 1795 points, a very good value that guarantees good performance in single-core applications.

Measured consumption was between 130 and 160W depending on the selected BIOS mode, which is very correct given the power developed.

The real surprise comes from the integrated GPU. I'll admit I had doubts before my tests. Forty RDNA 3.5 compute units clocked at 2900 MHz seemed light compared to the monsters equipping dedicated graphics cards.

Then I launched Counter-Strike 2 in 1080p with high settings. The game ran at 85-90 fps with occasional drops to 60 in particularly busy scenes. Totally playable. Shadow of the Tomb Raider, admittedly aging but still demanding, displayed 127 fps in 1080p with everything maxed. Even pushing up to 4K, the game remained fluid at 40 fps before any graphics parameter optimization.

Recent titles obviously demand more compromises. Monster Hunter Wilds is at 54 fps in 1080p with ultra settings, but collapses to 26 fps in 4K. Final Fantasy XIV: Dawntrail displays a comfortable 114 fps in 1080p which drops to 48 FPS in 4K with high settings. Clair Obscur: Expedition 33 runs at 50-55 fps in 1080p on medium settings.

Basically, the MS-S1 Max is a viable 1080p/1440p gaming machine. For native 4K in recent AAA titles, a dedicated graphics card remains essential. But for an APU without a graphics card, performance far exceeds what we're used to.

Can you replace a real graphics card with this system? The answer depends on your usage. If you mainly play titles like CS2, Valorant, League of Legends, or Apex Legends, the Radeon 8060S does the job handily. These games favor fluidity over graphical beauty and run without problem at over 100 fps.

For recent AAA titles, the situation gets complicated. The GPU holds up in 1080p with compromises on parameters, but forget intensive ray tracing or native 4K. Compared to an RTX 4060 Ti, the Radeon 8060S lags 30 to 40% in raw performance.

Artificial intelligence is this system's selling point. The ability to allocate up to 96GB of VRAM to the GPU allows execution of language models (LLM) locally. I'll admit this is quite new for us, but we'll certainly get into the habit of running some LLMs on test machines to compare them.

On an RTX 4090 with its 24GB of VRAM, you're limited to models of 30 to 40 billion parameters maximum. The MS-S1 Max can load models of 120 billion parameters.

So why run an LLM locally? This opens unprecedented possibilities for privacy, development, and also sensitive professional applications. One can imagine having this machine accessible to a few developers or others, for example.

I spent an entire week playing with different models under Linux and Windows. Under Ubuntu Linux 22.04, installing the necessary tools requires some technical skills but remains documented. GPU/CPU memory allocation is finely managed via BIOS. Setting dedicated VRAM to 1GB suffices, the Linux kernel dynamically allocates the rest according to needs.

Performance with ROCm 7 reaches 37-38 tokens per second on GPT-OSS 120 billion parameters. This throughput allows quasi-real-time interaction for code assistance, document analysis, or content generation.

Windows has caught up to Linux in AI matters. How? ROCm 7 which now runs natively without WSL. Installation requires several steps: Git for Windows, UV to manage Python environments, then downloading PyTorch modules from AMD repositories. The process takes about an hour the first time.

Once configured, Comfy UI starts without problem and automatically detects the 50GB of available shared VRAM in default configuration. LM Studio also works very well with the choice between ROCm and Vulkan.

Historically, Vulkan offered better performance on AMD hardware for AI. With the recent launch of ROCm 7, the situation reverses. On GPT-OSS 120B, Vulkan generates about 36 tokens per second while ROCm reaches 37-38 tokens per second. Smaller models like Qwen 3 4B run around 25-26 tokens per second.

Image generation with Comfy UI works well. Stable Diffusion XL, a memory-hungry model, loads entirely on the GPU without problem. A 1024 x 1024 image generates in about 12 seconds depending on query complexity obviously. This is less than an RTX 4070 mobile, for example, which takes 8 seconds on the same task. But the Radeon 8060S iGPU honorably holds the comparison.

A 70 billion parameter model in 8-bit requires about 70GB of memory. Technically possible on the MS-S1 Max, but the system will use the SSD to manage extended context. Performance then drops dramatically. I therefore advise 13 to 30 billion parameter models.

The integrated NPU capable of 50 TOPS (126 TOPS combined CPU/GPU/NPU) remains underutilized currently. Most AI frameworks essentially use the GPU for its versatility. Some specific applications like Windows Copilot or Adobe Photoshop with its native AI functions take advantage of the NPU, but the ecosystem is limited.

Compared to an RTX 5090 or RTX 6000 Ada professional, the MS-S1 Max obviously loses in raw performance. A 4090 generates 80 to 100 tokens per second on the same models thanks to its superior computing power and 1 TB/s memory bandwidth. But the 4090 only has 24GB of VRAM. For large models exceeding this capacity, impossible to load them. The MS-S1 Max therefore sacrifices pure speed to offer memory capacity without equivalent in its price range.

Windows 11 Pro arrives preinstalled on the 2TB SSD without pre-installed software. Minisforum delivers a clean system with only the necessary drivers.

All components are recognized from first boot: 10 GbE network ports work, Wi-Fi 7 connects, USB4 v2 ports respond. AMD Adrenaline updates download automatically via the brand's software.

Linux is generally well supported despite some small issues. Ubuntu 22.04 and later versions immediately recognize the processor, memory, and Radeon 8060S GPU. RDNA 3.5 drivers are integrated into kernel 6.x. The problem comes from Realtek R8125 controllers of 10 GbE ports.

Under Ubuntu, 10 GbE ports work after installing the r8125-dkms package available in community repositories.

Why Linux? The AI experience under Linux surpasses Windows. Memory allocation GPU/CPU is managed with more finesse. Setting dedicated VRAM to 512 MB in BIOS suffices amply, the Linux kernel dynamically allocates the rest according to application needs. This avoids juggling with BIOS settings.

Finally, performance with ROCm under Ubuntu reaches 37-38 tokens per second on GPT-OSS 120B, very slightly superior to Windows.

Power consumption at idle pleasantly surprised me. The system draws only 13-16W in standby, a surprisingly low figure for these specs. In light CPU load during web browsing or word processing, consumption rises to 45-58W. Gaming makes the counter climb between 110 and 140W depending on scene intensity.

I only managed to exceed 160W during the most sadistic tests. The 320W power supply therefore has a comfortable reserve.

At rest, the external chassis temperature quickly rose to 55-60°C while internal sensors displayed only 40-50°C. Fans remained discreet. As soon as an intensive workload launched, fans ramped up and external temperature dropped to 30-35°C. This means heat pipes efficiently transport heat to the fins once airflow increases.

Sound levels vary considerably depending on usage. At rest in Balanced mode, the system emits a discreet whoosh around 41 dBA, barely perceptible in a quiet office. Under maximum load with 3DMark, noise climbs up to 53 dBA. The sound profile reminds me of continuous low hum rather than unpleasant high-pitched whistling.

Small downside, thermal management still seems perfectible at software level. The fan control algorithm could react more quickly to load variations. During my first tests, the system remained abnormally hot at rest before cooling brutally once launched. Minisforum has already published two BIOS updates since product release, proof the company actively works on optimization.

The Minisforum MS-S1 Max is listed at €2,599 (128GB RAM and 2TB SSD). A price that can surprise when compared to classic mini PCs at €400 or €800. But this figure must be placed in the current context of the memory market.

Since late 2024, RAM prices have climbed 30 to 40% in six months. A 128GB DDR5-6000 kit today costs between €450 and €550, versus €350 a year ago. LPDDR5X-8000 MHz used in the MS-S1 Max is even more expensive due to its limited production.

The Geekom A9 Mega also offers the Ryzen AI Max+ 395 with 128GB RAM in a similar format sold at a similar price to the Minisforum. But Geekom includes a third native video output and uses Intel rather than Realtek network controllers for 10 GbE ports. Geekom's cooling system seems slightly less dimensioned.

The Framework Desktop, sold at €2,199, is another option with the same APU. Framework bets everything on repairability and modularity with easily replaceable components. The company promises rigorous software monitoring and the brand benefits from an active community. The weak point is in the absence of an open PCIe x4 slot for expansions.

Finally, Mac Studio M2 Ultra and Mac mini M4 Pro are a radically different philosophy. The Mac mini M4 Pro starts at €1,649 with 24GB unified RAM, up to €2,869 for 64GB memory and 2TB storage. This configuration offers exceptional performance in applications optimized for macOS. But the 64GB RAM limit on M4 Pro blocks execution of large AI models. To exceed this barrier, you must switch to Mac Studio M2 Ultra which goes up to 192GB unified RAM. Price then climbs beyond €6,000 in this configuration.

Finally, there are the brand new Nvidia DGX Spark machines. They're marketed under Nvidia, Dell, Lenovo, Asus brands with 240W consumption and 1 petaflop performance with FP4 precision, and the famous 128GB unified memory. The interest of these machines is integration of an Nvidia GB10 "superchip" composed of a 20-core Nvidia Grace ARM processor and Nvidia Blackwell graphics chip.