Managing GPUs and accelerators for VPC

Configuring a virtual server instance with an NVIDIA GPU

1. Provision a virtual server instance by choosing an NVIDIA GPU profile in the Profile field. Stock and custom operating system images are supported.

2. Install the NVIDIA GPU driver for your virtual server instance's image and GPU profile. The following table describes minimum driver and CUDA software version levels for Linux and Windows operating systems. For more information, see NVIDIA's Download drivers page. For an overview of drivers for NVIDIA data center products, see NVIDIA Data Center Drivers.

   NVIDIA drivers and CUDA version for Linux
   GPUs and minimum NVIDIA drivers and CUDA versions

   GPU	NVIDIA driver	CUDA version
   A100	550	12.4
   L4	550	12.4
   L40s	550	12.4
   V100	535	12.2
   H100	550	12.4
   H200	570	12.8
   B300 Select availability	590	13.1

   NVIDIA drivers and CUDA version for Windows 2019, 2022, 2025
   GPUs and minimum NVIDIA drivers and CUDA versions

   GPU	NVIDIA driver	CUDA version
   A100	538	12.2
   L4	538	12.2
   L40s	538	12.2
   V100	535	12.2
   H100	N/A	N/A
   H200	N/A	N/A
   B300	N/A	N/A

   NVIDIA drivers and CUDA version for Windows 2016
   GPUs and minimum NVIDIA drivers and CUDA versions

   GPU	NVIDIA driver	CUDA version
   A100	529	12.0
   L4	529	12.0
   L40s	N/A	N/A
   V100	535	12.0
   H100	N/A	N/A
   H200	N/A	N/A
   B300	N/A	N/A

3. Install associated toolkit for your workload. Visit NVIDIA's CUDA toolkit downloads page.

For detailed instructions to complete Steps 2 and 3, other GPU tools, and examples, see How to Use V100-Based GPUs on IBM Cloud VPC.

For a Linux-focused guide on installing the NVIDIA drivers, see the NVIDIA Driver Installation Guide.

If you want to automate the installation of the drivers, you can use the User data section of the virtual server. By using the user data field, you can input a script that issues the commands to install the NVIDIA drivers.

Configuring a virtual server instance with an NVIDIA B300 GPU

NVIDIA HGX B300 accelerated virtual server profiles are available for select customers. Create a support case if you are interested in purchasing and using this offering.

When you use the B300 GPU profile, the guest operating system needs an update. If the guest OS is not modified, you might see errors such as, "NVRM: This PCI I/O region assigned to your NVIDIA device is invalid:"

To configure the guest OS for the B300 GPU:

1. These commands must be run as root. Sudo to root.

   sudo -i
   

2. Edit the grub configuration file. The following example uses vi.

   vi /etc/default/grub.d/50-cloudimg-settings.cfg
   

3. Add the following kernel parameters to the GRUB_CMDLINE_LINUX_DEFAULT line:

   pci=assign-busses pci=realloc pci=nocrs pci=big_root_window
   

   The updated line should look like this:

   # CLOUD_IMG: This file was created/modified by the Cloud Image build process
   GRUB_CMDLINE_LINUX_DEFAULT="nofb console=ttyS0 console=tty1 pci=assign-busses pci=realloc pci=nocrs pci=big_root_window"
   

4. Update grub to apply the changes.

   update-grub2
   

5. Restart the virtual server.

   reboot
   

Configuring a virtual server instance with an Intel Gaudi 3 AI Accelerator

1. Provision a virtual server instance by choosing the Intel® Gaudi® 3 AI Accelerator instance profile in the Profile field. Stock and custom operating system images are supported.
2. Install the Intel Gaudi 3 AI Accelerator software and drivers for your virtual server. To download the drivers, see Intel Gaudi Driver and Software Installation page.

Integrating drivers into a custom image from volume

1. Provision a virtual server instance with a GPU and install the drivers.
2. Create an image from the virtual server instance stock image boot volume. For more information, see Creating an image from a volume.
3. Repeat the Image from volume process to deploy across multiple instances.

Next steps

For more information, see the NVIDIA driver documentation.