The AMD Ryzen AI Max+ 395 APU is a highly-integrated system-on-a-chip (SoC) featuring a heterogeneous architecture critical for efficient AI workloads:

• Zen 5 CPU Cores: Provide general-purpose processing and host for the Ollama service.
• Radeon 8060S iGPU (RDNA 3.5): The primary accelerator for LLM inference, utilizing shared system memory as Unified Memory (VRAM).
• XDNA 2 Neural Processing Unit (NPU): Designed for high TOPS AI acceleration, with potential for hybrid execution (prefill/context ingestion on NPU, decoding on iGPU/CPU).

Before deploying Ollama, the base Arch Linux installation must have the correct drivers and utilities to fully expose the APU's capabilities, especially for Vulkan and unified memory management.

Ensure the system is running a recent kernel (e.g., $6.10+$ or later) for optimal Zen 5 and RDNA 3.5 support.

snippet.bash

# Update system and install a recent kernel if not already running
sudo pacman -Syu linux linux-headers

The Vulkan backend relies on the open-source Mesa stack via the RADV driver.

snippet.bash

# Install Mesa with Vulkan support for AMDGPU
sudo pacman -S mesa vulkan-radeon lib32-vulkan-radeon

While the objective is Vulkan, installing the ROCm stack is often necessary for complete AMD GPU compute support and may be leveraged by other frameworks or future Ollama features. The Ryzen AI Max+ 395 is based on the gfx1150 target, which has improved support in recent ROCm versions (e.g., $6.4+$).

snippet.bash

# Install essential ROCm packages
sudo pacman -S rocm-core
# Install rocm-hip-sdk if developing or using other tools
# sudo pacman -S rocm-hip-sdk

The iGPU uses shared system RAM (Unified Memory). For optimal LLM performance, a large dedicated memory pool for graphics (iGPU/VRAM) is essential.

• BIOS/UEFI Configuration: Set the UMA Frame Buffer Size (or equivalent) in the BIOS/UEFI to the maximum supported value (e.g., 16GB or more, depending on total system RAM). This is the most crucial step for allocating dedicated VRAM.
• Kernel Parameters (Optional): For fine-tuning, verify kernel parameters if necessary, but the BIOS setting is typically sufficient.

Ollama relies on an underlying LLM engine (typically a fork of llama.cpp). Recent Ollama releases (e.g., $0.12.6+$) include experimental support for Vulkan acceleration, which must be explicitly enabled via an environment variable.

The easiest method is using the Arch package manager or downloading the official binary.

snippet.bash

# Install Ollama from the Arch Linux repository (or AUR)
sudo pacman -S ollama
# OR
# curl -L https://ollama.com/download/install.sh | sh

To force Ollama to utilize the more performant Vulkan backend on the AMD iGPU, the OLLAMA_VULKAN environment variable must be set when running the service.

Modify the Ollama systemd service unit to include the required environment variable.

1. Create an override directory:

   bash sudo mkdir -p /etc/systemd/system/ollama.service.d/

2. Create the override file (/etc/systemd/system/ollama.service.d/vulkan_override.conf):

   ini [Service] Environment="OLLAMA_VULKAN=1" # Optional: Set the number of threads for the CPU fallback/host operations # Environment="OLLAMA_NUM_THREADS=16"

3. Reload the daemon and restart the service:

   bash sudo systemctl daemon-reload sudo systemctl restart ollama

Check the service logs for confirmation that Vulkan initialization was successful:

snippet.bash

sudo journalctl -u ollama -f

Look for messages indicating Vulkan/GGML initialization on the iGPU. Ollama may also log which accelerator is being used when a model is run.

Test a small model, ensuring the output indicates GPU/Vulkan usage. The number of layers offloaded (--gpu-layers) is often determined automatically by the available VRAM (shared RAM).

snippet.bash

ollama run llama3:8b
# After the model downloads, monitor system resource usage (e.g., with htop and radeontop)
# The prompt prefill phase will typically show high iGPU usage.

The Ryzen AI Max+ 395 introduces the XDNA 2 NPU, enabling true heterogeneous computing. While the Ollama (via llama.cpp) Vulkan backend accelerates the iGPU, direct, standardized, and easily-configured NPU acceleration support within Ollama on Linux is currently limited/experimental and often requires niche frameworks or ONNX models (e.g., for hybrid execution).

• iGPU (Vulkan): Offers the best performance for general-purpose LLM inference, especially for larger models ($8B+$). This is the focus of the Vulkan optimization.
• NPU (XDNA 2): Best leveraged by specific, often proprietary, toolchains (e.g., Lemonade, ONNXRuntime GenAI) for models converted into an optimized format (e.g., OGA/ONNX).

The optimal, long-term deployment strategy will involve splitting the LLM workload to maximize the strengths of each compute unit.

\<pre class=“mermaid”> graph LR subgraph Frontend A[User Prompt] –> B(Ollama Server/API) end B –> C{Workload Scheduler (Ollama)} C –> |Prompt/Context Prefill (High TTFT focus)| D(XDNA 2 NPU) C –> |Token Decoding (High TPS focus)| E(Radeon 8060S iGPU - Vulkan) D –> F[NPU Results] E –> G[iGPU Results] F & G –> H(Final Token Stream) H –> B \</pre>

**Figure: Conceptual Heterogeneous LLM Pipeline for AMD Ryzen AI Max+ 395**

### 3. NPU Exploration (Advanced)

For the skilled Linux user aiming for NPU utilization, direct integration requires bypassing Ollama for NPU-specific frameworks:

• ONNX-EP: Investigate using ONNX Runtime with the AMD EP for NPU-accelerated execution, typically requiring model conversion and separate serving.
• AMD Guides: Follow AMD's specific Linux guides for the Ryzen AI 300 series to enable NPU drivers and libraries, then integrate them with a custom LLM serving solution that can utilize them (e.g., a hand-built llama.cpp with NPU support or a dedicated NPU LLM tool).

Would you like a step-by-step guide on how to build Ollama from source on Arch Linux to ensure the Vulkan backend is compiled correctly?