Skip to content

MS-S1 MAX Server

Setup guide for the MS-S1 MAX mini-PC running Ubuntu Server with ZFS, KVM, and Docker.

AI-Authored, Selectively Verified

This site was initially generated by Claude (Anthropic) and is being progressively reviewed against this specific hardware (MS-S1 MAX, AMD Strix Halo, Ubuntu 26.04 LTS). The Ubuntu install path, the ZFS section, the AI/ROCm guidance, and the rebuild checklist have been manually reviewed; some operational sections remain unverified. Always review root/sudo commands before executing.

Project Goals

Build a clean, minimal Ubuntu Server host that:

  • Uses ZFS for all important data
  • Runs full desktop OSes (Windows 11, Linux) via KVM/QEMU with GPU passthrough
  • Runs services (Nextcloud, Plex, etc.) via Docker
  • Remains rebuildable, understandable, and low-maintenance

The Hardware

This guide is built around the Minisforum MS-S1 MAX, a mini-PC featuring AMD's Strix Halo APU - a chip that combines CPU and GPU on a single die with access to all system memory.

AMD Ryzen AI Max+ 395 (Strix Halo APU):
+--------------------------------------------------+
|                                                  |
|  +-------------+          +------------------+   |
|  |   Zen 5     |          |    RDNA 3.5      |   |
|  |   16 cores  |          |    40 CUs        |   |
|  |   32 threads|          |    (gfx1151)     |   |
|  +------+------+          +--------+---------+   |
|         |                          |             |
|         +------------+-------------+             |
|                      |                           |
|         +------------v-------------+             |
|         |   Unified Memory         |             |
|         |   128GB LPDDR5X-8000     |             |
|         |   Quad-channel, soldered |             |
|         |   ~256 GB/s peak         |             |
|         +--------------------------+             |
|                                                  |
+--------------------------------------------------+
Component Specification Why It Matters
CPU Ryzen AI Max+ 395 (Strix Halo, 16C / 32T) Fast prompt processing
GPU Radeon 8060S, RDNA 3.5 (40 CUs, gfx1151) Accelerated inference
RAM 128GB LPDDR5X-8000, quad-channel Fits 70B+ models; ~256 GB/s peak
Architecture Unified memory No VRAM bottleneck

Unlike discrete GPUs limited to 24GB VRAM, the APU shares all 128GB with both CPU and GPU, enabling models that won't fit on any consumer graphics card. The LPDDR5X-8000 quad-channel bus delivers roughly 3x the bandwidth of typical desktop DDR5, which is what makes integrated-GPU inference practical at this scale.

Hardware details Hardware architecture

  • Getting Started

    Hardware specs, architecture overview, and prerequisites

    Get started

  • Ubuntu Server

    Installation, post-install config, networking, and firewall

    Ubuntu setup

  • ZFS Storage

    Pool creation, datasets, and snapshot policies

    ZFS guide

  • Virtualization

    KVM setup, GPU passthrough, and Windows 11 VM

    VMs

  • Docker Services

    Nextcloud, Plex, and other containerized services

    Services

  • Operations

    Backup, recovery, and rebuild procedures

    Ops guide

  • AI & Local LLMs

    Run local LLMs with Ollama, llama.cpp, and AI coding tools

    AI guide

Design Philosophy

Principle Implementation
Host OS is boring Ubuntu Server LTS, no desktop, SSH-only
Data lives outside containers ZFS is the source of truth
VMs are first-class KVM/QEMU with GPU passthrough
Services are containerized Docker + Compose with bind mounts
Everything is recoverable Reinstall host without touching data

What This Project Avoids

  • ZFS on root
  • Desktop environment on the host
  • Manual iptables rules
  • Docker volumes for critical data
  • Nested virtualization
  • "All-in-one" hypervisor distros