• Introduction
• Preparation
• Assembly
• Setup
• Testing
• BOM
• License

This tutorial is for anyone aiming to build a high-performance AI server with 8 GPUs. Whether you're a researcher, developer, or enthusiast, you'll learn everything from hardware selection and assembly to system configuration and initial testing. Finish with a robust platform ready for demanding AI workloads.

See detailed steps

Tip: The default BIOS settings may not deliver optimal performance for multi-GPU workloads. Adjust these parameters for best results:

• PCIe Settings
  🚨📢🔔⚠️ Set all PCIe slots to the highest supported speed (Gen4/Gen5) and configure bifurcation for your GPUs.
  

  Advanced -> Chipset Configuration -> PCIE link width -> set MCIO2/1, MCIO4/3, MCIO6/5, MCIO8/7, MCIO12/11, MCIO14/13, MCIO16/15, MCIO18/17 to x16
  

• Above 4G Decoding
  🚨📢🔔⚠️ Enable "Above 4G Decoding" to address large GPU memory.
  

  May be enabled by default
  

• Resizable BAR
  🚨📢🔔⚠️ Activate "Resizable BAR" for improved CPU-GPU data transfer.
  

  Advanced -> PCI Subsystems Settings -> Enable Re-size BAR support
  

• Power Management
  Disable unnecessary power-saving features (C-states, ASPM) that may throttle GPU performance.
  Optional

• Memory Configuration
  Set RAM to rated speed and enable XMP/DOCP profiles for max bandwidth.
  Optional

• Fan and Thermal Controls
  Adjust fan curves and thermal limits for optimal cooling.
  Optional

After saving changes, reboot and monitor GPU performance and stability.

References:

• Motherboard User Manual 💻🖱️
• BMC Documents 🤖

Boot with WinPE from USB to verify hardware, or install Linux, NVIDIA drivers, and check with nvtop. Once confirmed, install your OS and start your AI work.

• 𝄜 Bill of Materials

This project is open source under the MIT License.