Skip to main content

Preparing Bare Metal Nodes for A100 GPU Workloads

Deploying NVIDIA A100 GPUs in a private cloud environment requires careful preparation at the node level. Proper configuration ensures the system can fully leverage the capabilities of the hardware while enabling advanced features such as virtualization and multi-instance GPU partitioning. This article outlines the steps required to prepare a bare metal node for A100 GPU workloads.

Kernel and System Requirements​

For compatibility with the A100 GPU and related technologies like SR-IOV and MIG, the host system must meet specific kernel and BIOS requirements. OpenMetal recommends using a Linux kernel version 5.4 or higher. This ensures adequate support for GPU pass-through, I/O virtualization, and the necessary PCIe features.

At the kernel level, enable the following parameters for optimal operation:

GRUB_CMDLINE_LINUX_DEFAULT="quiet intel_iommu=on iommu=pt kvm.ignore_msrs=1"

The intel_iommu=on flag activates the IOMMU, allowing the system to manage device isolation and virtualization. The iommu=pt option ensures performance is not degraded by unnecessary address translations. For AMD systems, replace the IOMMU option accordingly:

GRUB_CMDLINE_LINUX_DEFAULT="quiet amd_iommu=on iommu=pt kvm.ignore_msrs=1"

After modifying the grub configuration, apply the changes:

sudo update-grub

Reboot the system to ensure the new parameters are loaded.

PCI Passthrough and SR-IOV Enablement​

SR-IOV (Single Root I/O Virtualization) is essential for deploying virtual GPU (vGPU) instances. Enable the necessary PCIe parameters to allow the GPU to present multiple virtual functions (VFs):

echo 1 > /sys/bus/pci/devices/0000:81:00.0/sriov_numvfs

Replace 0000:81:00.0 with the PCI address of the target GPU. The number of virtual functions available depends on the GPU model and configuration.

Use the following command to validate the presence of the GPU and confirm SR-IOV support:

lspci -nn | grep -i nvidia

BIOS Configuration for Virtualization​

Access the system BIOS or UEFI firmware and enable the following settings:

  • Intel VT-x (or AMD-V)
  • VT-d (or AMD IOMMU)
  • SR-IOV Support
  • Above 4G Decoding

These settings allow the hardware to support virtualization, large memory mappings, and multiple PCI devices—a requirement for multi-GPU workloads and partitioning.

Operating System Preparation​

Update the system and install basic tools required for NVIDIA driver installation and GPU management:

sudo apt update
sudo apt install -y build-essential dkms pciutils

Ensure that secure boot is disabled if kernel module signing is not configured, as the NVIDIA driver installation requires building and loading kernel modules.

Verifying PCIe Topology and GPU Recognition​

Use the lspci command to verify that the GPU is detected and mapped correctly:

lspci -nn | grep -i nvidia

For detailed PCIe hierarchy and bandwidth validation, run:

lspci -tv

This helps confirm the GPU is connected to a suitable PCIe slot providing the required bandwidth for compute workloads.

Next Steps​

With the node configured, the next phase involves installing the NVIDIA driver and setting up the environment for vGPU or MIG operation. Proper installation of the NVIDIA software stack is critical for managing the A100 GPU and enabling advanced features like multi-instance partitioning.

OpenMetal provides private clouds built to support GPU-based workloads. These environments ensure dedicated hardware access, predictable performance, and isolation for AI/ML and high-performance computing tasks.