Node A — Beelink SER9

CPUAMD Ryzen AI 9 HX 370
CPU ArchitectureZen 5
Cores / Threads12 / 24
Max Clock5.1 GHz
RAM32 GB LPDDR5X
SSD1 TB PCIe 4
Integrated GPURadeon 890M
AI NPUYes (~50 TOPS)
USB4Yes
Wi-FiWi-Fi 6
OSWindows 11
NetworkTailscale node ai-lab-b (100.115.162.13)

No discrete GPU — the Radeon 890M integrated graphics plus the onboard NPU are what's available for acceleration. Comfortable running 7B–9B models; 13B–14B models like Qwen2.5-Coder run fine at 4-bit quantization but slower; 30B+ would load but crawl without a dGPU to offload to.

Node B — GMKtec M7 Ultra

CPUAMD Ryzen 7 Pro 6850U
CPU ArchitectureZen 3+
Cores / Threads8 / 16
Max Clock4.7 GHz
RAM32 GB DDR5
SSD1 TB PCIe 4
Integrated GPURadeon 680M
AI NPUNone
USB4Yes
Wi-FiWi-Fi 6
OSLinux
RuntimeOllama
Networkreachable at 10.0.0.75:11434 on Node B's LAN

No AI NPU on this board, so model acceleration relies on the CPU and the Radeon 680M integrated graphics.

From the build

Spec comparison: GMKtec M7 Ultra vs Beelink SER9
The source spec sheet for the tables above

Side by side, the gap that matters most for local AI is the NPU row — one node has ~50 TOPS to lean on, the other doesn't.

For the network diagrams showing how these two boxes connect, see the Local LLM Field Guide, Chapter 10.