UPDATE Apr 13th 2026: The article is updated. I have since upgraded to Ubuntu 26 LTS Beta and gotten big increased in speeds and stability on many of the LLM models that I use esp those larger models in llama.cpp gguf quant. I’ve updated Grub variables and other details in the article. I maxed out RAM for LLMs here around 110GB model for Nemotron-Super-120B Q6_K gguf MOE model at 65k context. Some are claiming

Update May 6th 2026:
Ubuntu 26 LTS fixes (non -beta like previous update) , and llama-server with Qwen3.6 and more fixes. Not going to update this article again, but just know things are getting better and always changing at this point. Looks like at this date the post is still viable, and settings have “settled” but you should always double check.

The Post:

I picked up the MinisForum MS-S1 MAX AI on curiosity impulse, prompted by news of incoming RAM/storage price hikes (more about that later ) . Researching for weeks I found Windows caps iGPU RAM allocation at 96GB , Linux users can use kernel tweaks for whatever RAM they want for inference. I chose 124GB , leaving 4GB for headless “minimal install” Ubuntu Linux. I installed 10TB of SSD storage for storing weights.

I came into AI as a skeptic, but as someone who’d spent months on arXiv whitepapers. The more I read, the more convinced I became that genuinely useful AI beyond research and coding agents was closer than realized. The AI puzzle was heating up with diversity of thought, LLM-as-Judge, and multi-agent systems, and agents. I couldn’t resist playing with it.

I wanted more room to experiment, and this machine cleared that bar while being cost effective enough to act as a NAS server (9-12w idle usage and 10GB ethernet). I also wanted to “vote” on the future of AI taking a new direction with APU’s, a trend that kicked off in full effect with the Strix Halo with every major tech company bringing or announcing their take on the concept soon after. Later, when the AI compute market leap frogs I can jump for a “real” upgrade.

Btw, They where not lying about those prices hikes. Look where we are at just a few months later.

Item	Price – Nov 2025	Price – Apr 2026	Notes
MinisForum MS-S1 Max Ai	~$2,500	~$4,000	Sold out. Must buy 3rd party
Nvidia DGX Spark (comparison)	~$4,500 (at time of purchase)	—	Paid ~$2,000 less by choosing the mini PC
Samsung 990 PRO NVMe (x2)	$299.99 ea (Nov 2025 sale)	$750–$999 ea	Purchased 2 drives

The Ryzen AI MAX+ 395 uses a unified memory architecture where the integrated GPU can access up to 128GB RAM with Linux. That’s a fundamentally different proposition than a discrete GPU with a fixed VRAM ceiling. Dense models up to 70B parameters run just not fast though still can output multiples of human typing speed for planning / reason while lots of use can be gathered from MOE models of Q4 quant size up to 120B (200B or higher dropping quants below Q4). I was to experiment more with multiple model techniques like diversity of thought , LLM as judge/jury, and other types of multi-agent systems (model based) to see how far a system like this could be pushed with competing models biases even if “slow” by AI Chat standards.

I threw in two Samsung 990 Pro 4TB NVMe drives, giving me roughly 10TB total for model weights, RAG data, and training sets. The stock drive was moved to the PCIe 4.0 x1 slot and mounted as /mnt/slow_storage. One of the new drives sits in the stock PCIe 4.0 x4 slot, and the second runs via an adapter in the PCIe 4.0 x16 slot (operating at x4 speeds), mounted as /mnt/storage.

My Installation Notes:

• Boot into stock Windows and grab your OS key via PowerShell before doing anything else. if you ever need to return the machine to stock you have what you need to remove Windows completely doing this. This is where you partition for dual boot if you are going that route.
• Moved the stock NVMe to the PCIe 4.0 x1 slot (mounted at /mnt/slow_storage), installed my new 990 PRO Samsung 4TB drives in the x4 and x16 slots. PCI 4.0 x16 slot needs adapter ( SABRENT NVMe M.2 SSD TO PCle X16/XB./XA CARD WITH ALUMINUM HEAT SINK), runs at PCI 4.0 x4
• Create mount points mkdir -p /mnt/slow_storage /mnt/storage . Used ext4 on the drives via Gparted. get UUID for the drives, add to /etc/fstab file and “mount -a” to mount them
• Installed Ubuntu w/ “Minimal Install”. Focus is a headless AI system, not a desktop (more RAM pls).Note: the Ubuntu installer did not wipe the storage drive cleanly. I had to boot a GParted USB to sort the partition table. Disable Secure Boot in BIOS to get through the initial install (you can re-enable it later with some work).

BIOS Settings for more RAM for LLMs:

• Wake on LAN: Off ( Can configure on , needs software config on Linux. Off due to Ubuntu 24 reports online. That was later fixed. Research this if you need to switch on as some linux configuration needed at the time of this writing)
• ERP: On
• UMA Buffer Size: 1GB (the kernel parameters below handle the rest dynamically)
• Secure Boot off ( for getting Ubuntu installed. Can configure “on” with extra steps if needed )

GRUB / Kernel Parameters for Unified Memory (Maximizing Unified RAM ):

These three parameters are what unlock the real potential of the Strix Halo platform for AI:

NOTE: Seems these settings auto detected and installed in Ubuntu 26 version. This is for Ubuntu 24 LTS

• iommu=pt : Sets IOMMU to pass-through mode, bypassing translation overhead for GPU memory access. Keeps latency low across both the iGPU and any RDMA network cards. (You might want to avoid this if running Proxmox for your Propaganda mobile proxy social media bots)
• amdgpu.gttsize=126976 : Raises the GPU’s Graphics Translation Table size to ~124GB, overriding the conservative driver default and allowing the iGPU to address nearly all of your system RAM.
• ttm.pages_limit=32505856 : Tells the Translation Table Manager it can pin up to ~124GB of system pages in high-speed RAM, so the GPU isn’t constantly evicting and re-pinning memory during inference.
• UPDATE: the amdttm var are for updating for Ubuntu 26 LTS as things have changes since I started this on Ubuntu 24 LTS

GRUB_CMDLINE_LINUX_DEFAULT="quiet splash amd_iommu=off ttm.pages_limit=32505856 ttm.page_pool_size=32505856 amdttm.pages_limit=32505856 amdttm.page_pool_size=32505856"

Add this line below to /etc/default/grub then run “update-grub” at the cli

---

(Updated Apr 2026) Kernel 6.18.7 Why It Matters for LLM Workloads.

UPDATE: Remember, My system has since been updated to Ubuntu 26 LTS, Kernel 7.0

Once the base system was stable, I moved to the 6.18.7 mainline kernel via the Ubuntu Mainline Kernel Installer. There are improvements for Strix Halo AI workloads that shipped across the 6.14–6.18 kernel range: First noticeable effect for me was much larger models could run easily.

Better AMDGPU / RDNA 3.5 Driver ,TTM Memory Management ,AMDKFD Compute Under-the-Hood GPU Improvements (6.14+) AMD pushed targeted fixes for the Strix Halo’s iGPU: hung compute jobs recover without a reboot, large memory allocations stay pinned more reliably, and the ROCm compute stack got scheduling improvements all of which directly benefit llama.cpp and vLLM performance.

llama.cpp: On 6.18.7, You’re less likely to hit mysterious kernel-level eviction events mid-inference when serving a 120B models for extended periods.

vLLM: vLLM’s multi-node clustering support for Strix Halo which routes tensor-parallel inference traffic over RDMA

---

For FAST inference setup without installing stuff on baremetal: For containerized, ready-to-run environments covering llama.cpp, vLLM, ComfyUI, and fine-tuning on Strix Halo, check out Donato Capitella’s Strix Halo Toolboxes the fastest way to get AI workloads running correctly without breaking your host system. If you’re interested in actual fine-tuning (LoRA, QLoRA, full FT) on this hardware, his LLM Fine-Tuning Toolbox covers that too.

The GRUB parameters, kernel recommendations, and toolbox versions here were current as of early 2026. The underlying hardware hasn’t changed, but the software stack keeps getting better and updates are constant at this point. As you already know this entire industry is quickly moving forward and changing daily. Keep your eye on Donatos Youtube channel and AMD’s updates going forward.