Minisforum MS-R1 Arm Mini PC Review: Addressing Homelab Potential, GPU Compatibility, and Power Efficiency Challenges

Summary:

The Minisforum MS-R1 is an Arm mini PC with a 12-core Cix CPU, Mali G720 iGPU, full-size PCIe slot, 64GB RAM, NVMe storage, and dual 10 Gbps networking.

While hardware design and expandability are excellent, the system faces significant challenges, particularly high idle power draw (17W), which is comparable to or worse than many Intel/AMD systems and other Arm devices.

Comparison of Idle Power Draw for various Arm mini PCs, showing the Minisforum MS-R1 at 17W [ 00:03:34 ]
CPU performance is decent, outperforming Raspberry Pi 5 but falling short of Apple's M1. Integrated GPU performance is better than a Raspberry Pi but has limited Vulkan support out-of-the-box.

Geekbench 6 scores comparing Minisforum MS-R1 with other Arm systems, showing its position relative to Raspberry Pi 5 and Mac Mini M1 [ 00:02:34 ]
Initial attempts to install an NVIDIA RTX A2000 dGPU faced driver issues on Minisforum's custom Debian image. However, installing a generic Ubuntu OS automatically provided full GPU acceleration, drastically improving AI inference (71 tokens/s vs 12 on CPU) and graphics benchmarks.

AI Benchmarks showing significant improvement with NVIDIA A2000 GPU (71.36 tg/s) over MS-R1 CPU (12.12 tg/s) for Llama models [ 00:08:21 ]
The system has homelab potential with its expansion options and published guides for Proxmox/Jellyfin, but its high power consumption and initial software/driver complexities make it a poor value proposition for its $400-$500 price compared to more efficient Intel, AMD, or Apple alternatives, unless specifically targeting Arm development.

Hardware overview [00:00:00]

The Minisforum MS-R1 is presented as a new Arm workstation PC, featuring a robust hardware design but with underlying issues.

Initial impressions
- The device appears well-designed and quiet, using a 19V power adapter.
- It's lauded as having some of the nicest hardware outside of Apple's ecosystem.
Key specifications
- Features a 12-core Arm CPU with a Mali G720 iGPU.
- Includes a full-size PCIe slot for expansion.
- Comes with NVMe storage.
- Supports Wi-Fi 6E.
  
  A sticker showing key specifications of the Minisforum MS-R1, including its CIX CP8180 processor, Arm Immortalis-G720 MC10 GPU, 64GB memory, and 1TB storage [ 00:00:49 ]
Port configuration
- Offers 9 USB ports, including two Type-C with DisplayPort.
- Provides HDMI and dual 10 Gbps network jacks.
- Includes an audio combo jack.
  
  Detailed diagram showing the front and back ports of the Minisforum MS-R1, including USB, HDMI, and dual 10Gbps network jacks [ 00:01:00 ]
Internal design and expandability
- The chassis is easily opened by pulling a lever, revealing internal components.
  
  An inside view of the Minisforum MS-R1 chassis, showing the internal layout and components [ 00:00:10 ]
- The package includes adapters for U.2 or a second M.2 drive, which can replace the Wi-Fi card for additional expansion.

Cix's 12-core Arm SoC [00:01:47]

The integrated GPU (iGPU) and CPU performance of the Cix 12-core Arm SoC show mixed results.

iGPU performance
- Using Minisforum's default Debian 12 installation, GLMark2 ran well, scoring significantly higher than a Raspberry Pi (6322 vs 1935).
  
  A bar chart comparing GLMark2 scores, showing the Minisforum MS-R1 significantly outperforming the Raspberry Pi 5 [ 00:02:00 ]
- Vulkan support was problematic, with vulkaninfo crashing and VKMark not running initially.
- GravityMark did run, showing the G720 iGPU performance is comparable to Qualcomm's Adreno 750 (Snapdragon 8cx Gen 3).
  
  A snippet from the GravityMark leaderboard showing the Adreno 750's performance relative to other GPUs [ 00:02:13 ]
CPU benchmarks
- Geekbench 6 scores were 1336 for single-core and 6773 for multi-core, beating SBCs like Raspberry Pi 5 and Radxa Orion O6.
- However, these scores are still considerably lower than Apple's 4-year-old M1 chip.
  
  A bar chart comparing Geekbench 6 scores (single-core and multi-core) for various Arm systems, including Minisforum MS-R1, Raspberry Pi 5, Radxa Orion O6, Windows Dev Kit 2023, and Mac Mini M1 [ 00:02:34 ]
- In real-world use under Linux, the system felt fast, capable of handling 4K YouTube video playback while multitasking.

Deeper problems [00:02:57]

The Minisforum MS-R1 exhibits significant power consumption issues, which is unusual for an Arm-based system.

Idle power draw
- The idle power draw of the MS-R1 is high at 17W, which is considerably more than the Raspberry Pi 5 (2.8W), Orange Pi 5 (1W), and Mac mini M4 (4.1W).
  
  A bar chart illustrating the idle power draw (in Watts) of various Arm mini PCs, where lower is better, showing the Minisforum MS-R1 consuming 17W [ 00:03:34 ]
- This is also higher than some modern Intel (GMKtec G3 Plus: 9.3W, LattePanda Mu: 5.5W) and AMD (Framework Desktop: 12.5W) mini PCs.
  
  A bar chart comparing idle power draw for various mini PCs, including Intel, AMD, and the Minisforum MS-R1, highlighting its 17W consumption [ 00:03:51 ]
Theories on high power consumption
- The high idle power draw is hypothesized to be due to the CPU core layout, requiring the chip to be powered up significantly for stability and efficient core-to-core memory access.
  
  A heatmap showing inter-core one-way data latency between CPU cores, illustrating latency variations [ 00:03:14 ]
- Minisforum and Cix are reportedly working on optimizing power draw.
Memory access speed
- Despite power concerns, memory access speeds (memcpy and memset) are relatively fast compared to other SBCs, though still below the Mac mini M4.
  
  A bar chart comparing memory access speeds (MB/sec) for various mini PCs, showing Minisforum MS-R1 with high speeds in both memcpy and memset tests [ 00:03:19 ]
Overall efficiency
- The HPL Efficiency benchmark (Gflops/W) shows the MS-R1 at 1.08 Gflops/W, which is less efficient than a Raspberry Pi 5 (3.2 Gflops/W) and significantly behind the Mac mini M4 (7.57 Gflops/W).
- This challenges the common perception of Arm systems being inherently more efficient.

10 Gbps Networking and AI inference [00:03:55]

The MS-R1 offers strong networking capabilities but has mixed results in CPU-based AI inference.

Networking performance
- The onboard NICs deliver full 10 Gbps speeds.
  
  Text overlay indicating Realtek RTL8127 x2 providing 9.40 Gbps networking speeds [ 00:03:58 ]
- Built-in Wi-Fi 6E achieved gigabit speeds on the network.
AI inference on CPU
- With 64GB of RAM, the system has potential for local AI inference on the CPU.
- Benchmarks using Llama models showed the MS-R1 underperforming the older Radxa Orion O6 board, despite using the same CPU, with scores ranging from 3.91 to 12.12 tokens per second.
  
  A bar chart comparing CPU-based AI inference benchmarks (tokens generated per second) for various systems, showing the Minisforum MS-R1 performance across different Llama models [ 00:04:13 ]

Installing a dGPU (RTX A2000) [00:04:55]

Adding a dedicated GPU (dGPU) to the MS-R1 for enhanced performance proved to be a challenging but ultimately successful endeavor with the right OS.

GPU selection and physical fit
- A half-height, single-slot NVIDIA RTX A2000 laptop GPU with 8GB VRAM (costing ~$250) was chosen to fit the tiny PCIe slot.
  
  The NVIDIA RTX A2000 SFF GPU shown, highlighted with its cost of about $250 [ 00:05:18 ]
- Minisforum included ventilation holes on the top of the chassis to aid cooling for such an installation.
- Installation was easy but finicky, with a foam spacer used to prevent shorts.
Driver issues on Debian
- After installation, the GPU was detected by lspci on Minisforum's default Debian image.
  
  Terminal output showing `nvidia-detect` successfully identifying the NVIDIA RTX A2000 Mobile GPU on the Debian system [ 00:05:42 ]
- However, attempts to install NVIDIA drivers (both direct downloads and Debian packages) failed, likely due to the Debian image being cross-compiled, making kernel module building difficult.
  
  Terminal output showing an error during the installation of NVIDIA drivers on the Debian system, indicating dependency problems [ 00:05:48 ]

A BIOS detour and AC Power Loss bug [00:06:15]

While attempting to resolve software issues, the reviewer explored the BIOS and found a bug with the AC Power Loss setting.

BIOS features
- The BIOS is quite comprehensive, offering settings for USB, RAM, and Power On behavior.
- Many settings are labeled "Beta," indicating ongoing development.
  
  A screenshot of the BIOS "System Manager" showing "Beta2" version and build date, with various configuration options below [ 00:06:23 ]
AC Power Loss bug
- The BIOS setting for "Restore AC Power Loss" (to turn on when power is restored) did not function as expected.
  
  A screenshot of the BIOS setting for "Restore AC Power Loss" with "Power Off" and "Power On" options, which was found not to work via software [ 00:06:32 ]
- It was discovered that a physical hardware switch on the motherboard, not the BIOS setting, controls the AC power loss behavior.
  
  A close-up of the motherboard showing a physical toggle switch labeled "AC LOSS ON / OFF" [ 00:06:43 ]

Trying an Intel GPU (A310 ECO) [00:06:50]

An alternative dGPU, the Intel Arc A310 ECO, was also tested but failed to function.

Installation and detection
- An Intel Arc A310 ECO with 4GB VRAM was installed, fitting physically better than the A2000.
- Upon boot, the fan on the A310 spun up, but lspci could not detect the card, indicating a signaling issue.
- This problem has been encountered with the same card on a Raspberry Pi.

Installing Ubuntu and getting GPU acceleration [00:07:20]

Switching to Ubuntu resolved the dGPU driver issues and unlocked significant performance improvements.

Easy Ubuntu installation
- Installing Ubuntu from an Arm ISO on a USB drive was straightforward.
- While the internal HDMI output froze during boot, switching the monitor to the A2000's HDMI port revealed a working desktop.
NVIDIA driver success
- Ubuntu automatically loaded the correct NVIDIA driver.
- CUDA could be installed, and nvidia-smi confirmed proper driver and GPU detection.
  
  Terminal output showing `nvidia-smi` successfully detecting the NVIDIA RTX A2000 Laptop GPU, its driver version, and CUDA version on Ubuntu [ 00:07:52 ]
Performance improvements with dGPU
- VKMark ran smoothly, scoring over 10,000 with the A2000, significantly higher than the iGPU's 3,037.
- AI benchmarks for Llama 3.1:3B model increased from 12 tokens/second (CPU) to 71 tokens/second (A2000).
- Efficiency also improved, reaching 0.76 tokens/watt on the GPU compared to 0.34 on the CPU.
- Larger 8-billion parameter models saw inference jump from 5.78 to 34.52 tokens/second.
  
  A bar chart comparing AI benchmarks (tokens generated per second) for the MS-R1 CPU versus the MS-R1 with an NVIDIA A2000 GPU across different Llama models [ 00:08:21 ]
- GravityMark also ran much smoother with the A2000, scoring 16,679 compared to the iGPU's 3,037.
  
  Benchmark results for GravityMark 1.89 Vulkan, showing a score of 16,679 with the NVIDIA RTX A2000 Laptop GPU [ 00:08:36 ]

Arm Homelab in a box? [00:08:40]

The Minisforum MS-R1 has potential as a homelab box due to its expansion capabilities but is hindered by its price-to-performance ratio and software immaturity.

Homelab suitability
- The full PCIe slot, coupled with U.2 and NVMe adapters, makes it highly expandable for storage and other cards.
  
  The Minisforum MS-R1 with its chassis partly slid open, revealing the internal components and PCIe slot [ 00:08:48 ]
- Minisforum provides guides for installing Proxmox and Jellyfin with hardware passthrough and iGPU acceleration, indicating its intended use as a homelab device.
- Its expansion options surpass other Arm machines, including the Mac mini.
Value proposition
- As an Arm desktop, it performs better than most SBCs.
- However, at $400-$500, it offers poor value compared to more efficient and mature Intel or AMD mini PCs, or even the Apple M1/M4 Mac mini (which offers superior performance but lacks bare-metal Linux support).
Current state and future outlook
- The MS-R1 is considered "half-baked" due to power consumption issues and driver complexities on generic Linux distributions.
- Questions remain about potential firmware fixes, mainline driver integration for all Linux distros, and NVIDIA Windows on Arm driver support.