Independent guide · methods and evidence
AMD Strix Halo Setup: BIOS, UMA, IOMMU, Ubuntu and Local LLMs
Set up a Strix Halo AI PC with the measured Ubuntu and Vulkan/RADV route, explicit Ollama versions, memory guidance and reproducibility checks.
Evidence reviewed .
Before copying commands: the local chat path and acceptance checks are written for a 128GB system on Ubuntu 24.04. Keep Ollama and any optional browser UI on loopback, and read the version-specific limits of that UI below.
Fresh installations through setup.sh request Ollama 0.31.2. Existing installations are retained. Only the requested runtime is pinned; model tags, drivers and the rest of the host stack still need verification. See the full setup behavior.
September 19 functional update: the scoped runtime qualification records text (Qwen3.6), image (Qwen2.5-VL), executed tools (Devstral) and pinned Open WebUI on the existing 0.32.15 service after restart. Isolated 0.34.2 is useful but not default; official Qwen3.8 and full-reboot candidate acceptance remain open. The tested host's Ollama listener was LAN-reachable, not local-only. Released llama.cpp v0.4.1 passed bounded direct/server/HIP controls; this does not qualify every model, long-context shape or maximum-memory allocation.
This is the practical starting configuration measured by the independent
AMD Strix Halo Local LLM Guide.
The guide's own setup evidence covers Ryzen AI MAX+ 395 / Radeon 8060S
(gfx1151) with 128GB installed; community results on 96GB are separate
evidence. The primary first-party machine is a 128GB Beelink GTR9
Pro; BIOS labels, firmware, cooling and power modes can differ on other OEM
systems.
Setup reviewed: November 9, 2026 (review of the recorded configuration; no new installation or measurement). Beginner route text reviewed: October 14, 2026; the configuration evidence keeps its original dates. Exact benchmark claims remain canonical in the repository's structured data and raw evidence.
The setup script preserves administrator Ollama drop-ins and stops on unresolved environment conflicts or legacy custom GPU rules. Existing installations may therefore need manual review; writing configuration does not qualify an upgrade or prove GPU offload. See the manual configuration and access checks.
Running the current official dense Qwen model? Use the dedicated Qwen3.8 27B on Strix Halo route comparison for the measured Ollama path, context boundary, MTP/DFlash distinctions, and current community performance leads.
If setup is already failing, use the symptom-first Strix Halo troubleshooting page. To choose between measured routes and newer unmeasured artifacts, use the Strix Halo model hub.
What Is AMD Strix Halo?
AMD Strix Halo is the hardware family behind Ryzen AI MAX systems. The Ryzen AI MAX+ 395 combines 16 Zen 5 CPU cores with Radeon 8060S integrated graphics (40 RDNA 3.5 compute units) and configurations with up to 128GB of LPDDR5x unified memory. AMD's Ryzen AI Halo developer platform uses this processor with 128GB of memory; retail systems from different OEMs can expose different BIOS, power, cooling and firmware behavior. See AMD's official Ryzen AI Halo specifications.
Unified memory does not mean that applications should reserve all physical RAM as fixed VRAM. Linux, the model runtime and other processes still need memory. On the measured Vulkan/RADV path, a small fixed UMA reserve leaves memory visible to Linux while the integrated GPU accesses a much larger GTT-backed shared pool.
Start With One Local Chat Path
If you want to install, start with Get local AI running, step by step. That page is the complete beginner route: what you need, the installation, your first answer, a question with a known answer, the GPU check, the current Ollama loopback-listener check, and help at every step. Browser and second-device access have separate checks. A chat window in your browser is the natural next step for everyday use. This page explains the configuration behind the route; the copyable commands live on that one page to reduce duplication. GPU answers and the pinned browser interface passed an isolated hardware check on October 15; a complete fresh native installation remains unverified. The start route records the test boundaries.
32GB, 64GB or 96GB installed? This guide has no tested automatic profile for those capacities. Use the manual memory reference to understand each setting, then qualify a profile for your own installed memory, GPU budget, model and context. The same applies to 192GB: do not copy the 128GB limits. A candidate model in the RAM card does not change this setup boundary.
What Has Been Tested?
- Ubuntu 24.04, 128GB: the October 15 isolated check used Ollama 0.31.2 and the pinned WebUI 0.11.4 image, separate ports and a bind-mounted new data directory. GPU output, browser login/answer, closed signup, rejected anonymous requests and process restart passed.
- Still open: Docker installation, the exact default-port/named-volume command, a fresh native installation, host reboot, restore and access from a second device. The component tests above do not cover these checks.
- Windows: use Starting on Windows. It has its own community and runtime evidence; Linux boot values and results do not qualify Windows.
If boot, GPU placement or browser chat fails, start with the free troubleshooting checks. Still stuck? Setup Rescue is paid personal help for a 128GB system on Ubuntu 24.04; other configurations get a bounded diagnosis only, after an intake.
Short Answer
For a normal retail AMD Strix Halo local-AI setup:
- Use the measured Ubuntu 24.04 LTS route; select X11 only if a desktop tool needs it.
- Set BIOS UMA Frame Buffer Size to 512MB if available, or 2GB if that is the vendor BIOS minimum.
- Leave IOMMU enabled/default for the normal buyer path, NPU use, suspend, RDMA, VFIO, passthrough and clustering.
- The recorded 128GB Beelink profile uses
amdgpu.gttsize=131072 ttm.pages_limit=31457280 amdgpu.cwsr_enable=0.cwsr_enable=0disables compute wave save/restore (mid-wave compute preemption); it is the gfx1151 workaround from ROCm/ROCm#5590. The GTT/TTM values are limits, not allocated VRAM; do not apply them as a 96GB preset. 192GB-class systems (for example Ryzen AI Max+ PRO 495) are not qualified; do not reuse the 128GB values there. Preserve unrelated boot parameters and verify live values after reboot. - Use Mesa/RADV and verify the selected ICD. Record a power policy per campaign;
tuned accelerator-performanceis one historical reproduction profile, not a requirement for every routine check. - Start with Ollama on Vulkan/RADV. Move to direct
llama.cppfor controlled benchmarks and to the documented ROCm or server routes only when the workload requires them.
192GB-class and PRO 495 systems (checked 2026-09-30). setup.sh stops above
about 136GiB of visible RAM, and no 192GB-class profile is qualified here. Vendors
advertise up to 160GB of dedicated VRAM through the BIOS for their 192GB systems
(for example the
HP ZBook Ultra G3a page
and the Minisforum MS-S1 MAX page);
this guide's route uses a small fixed UMA reserve plus GTT. Neither approach has
been tested on 192GB here. On Linux the NPU of the Ryzen AI Max+ PRO 495 is
expected only with kernel 7.4 unless a patch is backported (Phoronix,
2026-09-30: the XDNA driver patch targets 7.4, with its merge window in late
October and a stable kernel around the end of 2026; a backport is possible but
uncertain). No 495 system has been tested here.
Do not treat amd_iommu=off as a universal recommendation. It is an optional
always-on desktop benchmark profile used by some measured first-party runs; it
disables NPU access and can break mobile suspend.
IOMMU notes (checked 2026-09-30). Phoronix (2026-09-29) tested AMD's PerfOpt,
an IOMMU bypass for iGPU memory access that is on by default in Linux 7.4 and can
be turned off with amdgpu.iommu_perfopt=0. On a Framework Desktop (Ryzen AI
MAX+ 395, 64GB) with Lemonade it measured 2 to 4%; the 18 to 23% in its headline
applies to smaller iGPUs and should not be carried over to Strix Halo. The patch
was queued but not yet merged. Whether PerfOpt removes the roughly 6% memory-read
gain recorded here for amd_iommu=off in one desktop test is not measured.
Separately, amd_iommu=off also turns off the IOMMU that the kernel's USB4 and
Thunderbolt documentation names as the protection against DMA by connected
devices (kernel docs).
Kernel (checked 2026-09-30). Ubuntu's 7.0.0-28 HWE kernel (the updates kernel from 2026-07-16 to 2026-08-17) carries an amdgpu memory-management regression from upstream 7.0.12 that was fixed in 7.0.13; Ubuntu has the fix from 7.0.0-29 (Launchpad #2158267 reports a roughly 42x ComfyUI SDXL slowdown; 7.0.0-34 was in noble-updates on 2026-09-22). The reporter of ROCm#6508 (open) attributes a KFD hang on 7.0.0-28 to that kernel and saw none on 7.0.0-34. Use the current HWE kernel. Any ROCm/HIP result in this guide that was measured on 7.0.0-28 may be affected; the effect on those numbers is unknown.
Which Configuration Should I Use?
| Your use case | UMA setting | IOMMU policy | Starting runtime |
|---|---|---|---|
| Normal local-AI buyer path | 512MB if available, otherwise the vendor minimum such as 2GB | Enabled/default | Ollama with Vulkan/RADV |
| Laptop, suspend or NPU use | Vendor-supported minimum | Enabled/default | Ollama or the documented NPU route |
| RDMA, VFIO, passthrough or clustering | Vendor-supported minimum | Enabled, with iommu=pt only when the workflow needs pass-through behavior |
The documented workload-specific route |
| Strict reproduction of the measured always-on desktop benchmark profile | 512MB on the primary Beelink system | Optional amd_iommu=off, with the NPU and suspend caveats understood |
Direct llama.cpp with Vulkan/RADV |
Why The UMA Setting Is Small
The fixed UMA frame buffer is not the total memory available to the integrated GPU on the measured Linux path. Vulkan/RADV can use GPU-accessible unified system memory through GTT. On the primary 128GB Beelink system, using the small fixed UMA reserve left almost the full memory pool visible to Linux while the iGPU could still access the large GTT-backed pool.
Some vendor BIOSes expose 2GB rather than 512MB as the minimum. Use the lowest supported vendor setting that still exposes the expected shared-memory pool; do not assume every OEM uses the same label or range.
Install The Working Buyer Path
Follow the complete beginner route.
Read its sections on the required system, backups, disk space and security
first; after that it gives you every command, what it should print and where
to stop, on one page. The script does not install Ubuntu and does not touch the
BIOS. It changes the boot settings (GRUB), GPU access and the Ollama service
configuration, adds the kisak Mesa PPA, runs a full apt upgrade -y and
downloads a model. If it fails halfway, the earlier changes stay in place; the
manual undo table
exists but is untested on hardware.
The automatic route is for Ubuntu 24.04 on the measured 128GB-class hardware, with 120–136GiB visible to Linux. Check your installed memory and CPU yourself: the script does not check the exact processor model, and a larger machine that reserves memory in the BIOS can pass its visible-memory check. Other memory sizes and operating systems need their own testing. A fresh run installs Ollama 0.31.2; if Ollama is already installed, that version is kept. Neither the revised installer nor a newer Ollama has been through a complete fresh-install test here.
Security: published advisories include the historical 0.31.2 version. Ollama has no password, so keep its API listening only on this computer, and never open it to the network to make a browser interface work. The start route describes the digest-pinned Open WebUI 0.11.4 candidate; it does not tell you to install the historical 0.10.2 image. In an isolated component check on October 15 it passed startup with new data, first admin/login, a visible answer, closed signup, anonymous-request rejection, loopback listeners and UI/runtime restart. That check did not cover a fresh native installation, host reboot or restore. The later October 18 channel-webhook advisory affects versions >=0.7.0,<0.12.2, including this pin; 0.12.2 is the listed fix. The issue requires enabled channels, a channel webhook, member write access and knowledge of the webhook ID. The advisory says instances with channels disabled, or without channel webhooks, are not affected. Keep these features off; a loopback-only listener does not resolve an issue that an authenticated user can trigger. A newer version needs its own acceptance checks before it replaces the pin. Read Security status of pinned components and the local AI security checklist.
Choose The Backend By Workload
| Goal | Start here | Reason |
|---|---|---|
| Local chat and Open WebUI | Ollama with Vulkan/RADV | Easiest measured buyer path |
| Reproducible generation benchmark | Direct llama.cpp with Vulkan/RADV |
Exact control over model, quant, build and command |
| Local API, batching or speculative decoding | llama-server |
Supports the documented server and MTP experiments |
| Higher concurrency or prompt-heavy work | The measured Lemonade or ROCm/HIP profile | These routes can suit batching and prompt processing better than the single-stream default |
| vLLM, long-context or advanced ROCm work | Use only the specific documented route | Support and performance remain workload- and version-dependent |
There is no single backend that wins every Strix Halo workload. Direct
llama-bench, Ollama API, server, MTP/speculative, concurrency and community
results are separate claim types in this project.
Browser interface: the start route's Open WebUI section is the single place for the new 0.11.4 candidate and its checks. The historical 0.10.2 result is not current installation advice. The limited October 15 pass does not establish current security or complete installation acceptance; see Open WebUI in SECURITY.md.
Frequently Asked Setup Questions
What BIOS UMA setting should I use on Strix Halo?
Use 512MB when the BIOS offers it, or the lowest vendor-supported setting such as 2GB. Then verify that Linux still sees the expected system-memory pool. A large fixed UMA reservation is not required for the measured Vulkan/RADV path.
Should I disable IOMMU on Strix Halo?
Not for the normal buyer setup. Leave IOMMU enabled or at the firmware default
for NPU use, suspend, RDMA, VFIO, passthrough and clustering. amd_iommu=off is
only an optional reproduction profile for an always-on desktop benchmark box.
Which Linux distribution should I use for a Strix Halo local LLM?
This guide's measured beginner runtime uses Ubuntu 24.04 LTS and Mesa/RADV; X11 is conditional on desktop-tool needs. The revised script's configuration checks do not replace fresh-install/upgrade hardware qualification. Other distributions can work, but they are not automatic substitutes for this exact measured path; compare their kernel, Mesa, Vulkan ICD and runtime versions against the evidence. Ubuntu 26.04 is not qualified by this guide (checked 2026-09-25; the 26.04.1 point release shipped 2026-08-27). Community 26.04 reports stay separate from first-party results.
What This Setup Does Not Guarantee
- It does not guarantee identical performance across Beelink, Framework, Corsair, GMKtec, Minisforum, Nimo or other Strix Halo systems.
- It does not turn a low-bit capacity pass into a model-quality recommendation.
- It does not make experimental ROCm, MTP, vLLM or NPU paths beginner defaults.
- It is not official AMD or OEM documentation or endorsement.
For cross-OEM differences, use the system evidence matrix.
Evidence And Source Of Truth
- Complete guide and setup script
- Concise canonical setup with detailed caveats
- Workload-specific best-known profiles
- Retail buyer-path validation
- Cross-OEM system evidence matrix
- Structured headline claim index
- Raw evidence
If a number appears in a post or AI answer but not in the linked structured data or raw evidence, treat it as unverified.