Hardware
- Beelink GTR9 Pro (BIOS GTRP110, 2025-12-18)
- AMD Ryzen AI MAX+ 395 (Strix Halo) — 16C/32T, max boost 3 GHz
- 128 GB DDR5-8000 unified memory (8× Micron 16 GB SODIMM)
- 2× NVMe SSDs: Crucial P310 2 TB (system) + Samsung S4LV008 4 TB (VM data, ZFS)
- Intel E610 dual-port 10GBASE-T NIC
Operating system
- Proxmox VE 9.1.9 (running on Debian 13 base)
- Kernel 7.0.0-3-pve (proxmox-kernel-7.0)
- AMD microcode amd64-microcode 3.20251202.1 kernel-loaded; BIOS-side patch B70001E
- ZFS for VM storage on the secondary NVMe
iGPU configuration (Radeon 8060S)
- BIOS setting Dedicated Graphics Memory = 64 GB (UMA frame buffer) — splits unified
memory into 64 GB iGPU VRAM + 62 GB system
- Vulkan stack via mesa-radv (in-tree, RADV GFX1151, Mesa 25.0.7)
- ROCm support enabled in kernel config (
HSA_AMD=y, HSA_AMD_SVM=y, HSA_AMD_P2P=y)
- Used for local LLM inference via Ollama (v0.20.3, Vulkan backend,
OLLAMA_VULKAN=1)
- Current production workload: GLM-OCR (0.9B vision model, 2 GB VRAM) for production document-OCR workflows, plus a smaller Gemma4 model loaded in parallel
- Two stability issues observed during evaluation that may interest your engineering team:
- Larger models (Gemma4 30B-class) crashed the host twice during testing — full system
reset, not just an Ollama process failure. We rolled back to smaller models and stayed
there. Likely Vulkan buffer pressure or VRAM allocation path on Strix Halo, but we
didn’t isolate the root cause beyond “30B-class models are unsafe on this iGPU today”.
- Vulkan buffer allocations >4096 context tokens hit an 8 GB max-buffer ceiling
(ErrorOutOfDeviceMemory). Workaround: cap OLLAMA_CONTEXT_LENGTH=4096.
- Both observations seem driver/backend-related rather than hardware. If a future BIOS or AMD
driver update addresses Strix Halo Vulkan stability for larger models, we’d be very
interested to retest — the 64 GB VRAM headroom is what attracted us to this platform.
Networking
- Port 1 — primary, currently 10 Gbps full-duplex to TP-Link Archer BE800 router (Port 3, 10G WAN/LAN configured as LAN). Driver: in-tree ixgbe.
- Port 2 — reserved for future direct 10 Gbps cross-box link to a planned secondary node
- Wake-on-LAN enabled
Notable BIOS settings (non-default)
- Above 4G Decoding: Enabled (required for E610 NIC detection)
- Re-Size BAR Support: Enabled (iGPU performance)
- SR-IOV Support: Enabled (planned VM NIC passthrough)
- TPM 2.0 / Security Device Support: Enabled — used by VMs for sealed-secrets workflows.
Stability note
The system has been stable in production since BIOS GTRP110 with kernel 7.0.0-3-pve. Earlier kernels (≤6.14) had less complete Strix Halo support; we recommend ≥6.16 for anyone running this hardware on Linux.
