Getting Started

Getting Started

This is the recommended first path for a new Intel XPU system.

1. Inventory The System

Before changing drivers or installing frameworks, collect:

• GPU model and count
• motherboard, CPU, RAM, power supply, and risers
• operating system and kernel
• BIOS settings for Above 4G decoding, ReBAR, IOMMU, PCIe speed, and Secure Boot
• current driver packages
• visible XPU devices from Level Zero, OpenCL, PyTorch, OpenVINO, and framework tools

Use:

• Linux: scripts/collect_xpu_snapshot.sh
• Windows: scripts/collect_xpu_snapshot.ps1

These scripts are read-only except for writing a local snapshot folder.

2. Install Official Drivers First

Start from the official driver path for your OS.

• Windows: Intel Graphics Windows DCH drivers
• Linux: Intel GPU Linux documentation for your distro and kernel

Do not mix random package repos until you have a known-good baseline.

3. Verify Device Visibility

A useful baseline sees the GPUs through several layers:

• lspci or Device Manager
• xpu-smi if available
• Level Zero tools such as ze_info
• OpenCL tools such as clinfo
• PyTorch XPU
• OpenVINO device query
• the target framework, such as vLLM or llama.cpp

If one layer sees the GPU and another does not, post both outputs.

4. Pick A First Workload

Choose a small, known workload first:

• OpenVINO model device query or sample
• PyTorch XPU tensor smoke test
• llama.cpp SYCL small model smoke test
• vLLM XPU smoke test with a model that fits comfortably

Do not start with a huge multi-GPU model and driver tuning all at once.

5. Record The Baseline

Post the baseline in Discussions:

• snapshot script output
• installed driver/runtime versions
• exact command
• pass/fail
• benchmark numbers if any
• what you want to improve next