GPU Virtual Machines

GPU VMs provide dedicated GPU resources for machine learning, AI workloads, and high-performance computing applications.

Overview

GPU VMs are created the same way as CPU VMs - by selecting a GPU-enabled resource profile and an OS that supports GPUs. The evroc compute service offers GPU resource profiles with NVIDIA L40S and B200 GPUs. For available GPU profiles and specifications, see Virtual Machine Resource Profiles.

Note: GPU VMs currently support Ubuntu 24.04 only.

Dedicated GPU allocation

Each GPU is dedicated to a single VM rather than timeshared across multiple VMs. This ensures consistent performance and eliminates resource contention.

High-performance local disk

GPU VMs include high-performance local NVMe SSD storage at no additional cost. The storage capacity scales with the resource profile, ranging from 3.8 TB to 32 TB. This local disk is ephemeral. It cannot be transferred or mounted to another VM and data will be lost if:

  • The VM's host fails
  • The VM is stopped through the API, for example by updating with --running=false in the CLI.
  • The VM is deleted

Data is not lost if you reboot the VM from within the guest, for example by typing sudo reboot at the command line.

Default VM initialization

GPU VMs come preinstalled with:

  • NVIDIA open-source GPU drivers (nvidia-open)
  • NVIDIA container toolkit

This initialization requires outbound network access to:

  • https://developer.download.nvidia.com/
  • https://nvidia.github.io/
  • Default APT repository URLs

If using non-default security groups, configure them to allow this outbound traffic.

Overriding default initialization

Override the default initialization using custom cloud-init userdata. See Using custom cloud-init userdata for details.

Quotas

GPU VMs are subject to separate quotas from CPU VMs. See Service Definition for quota details.