Backend.AI is an open source cloud resource management platform, which makes it easy to utilize virtualized compute resource clusters in a cloud or on-premises environment. The container-based GPU virtualization technology of Backend.AI supports the efficient use of GPUs by flexibly dividing one physical GPU, so that multiple users can use it at the same time.

Backend.AI offers a variety of performance-driven optimizations for machine learning and high-performance computing clusters, along with management and research features to support a diverse users, including researchers, administrators, and DevOps. The Enterprise Edition adds support for multi-domain management, a dedicated Control-Panel for superadmins, and the GPU virtualization plug-in.

A GUI client package is provided to easily take advantage of the features supported by the Backend.AI server. Backend.AI Web-UI is a GUI client in the form of a web service or stand-alone app. It provides a convenient graphical interface for accessing the Backend.AI server to utilize computing resources and manage its environment. Backend.AI provides pre-made images which enable immediate creation of compute sessions without any need of installing separate programs. Most tasks can be done with mouse clicks and brief typing, which achieves more intuitive use.

Key Concepts

Diagram explaining key concepts
  • User: The user is a person who connects to Backend.AI and performs work. Users are divided into normal users, domain admins, and superadmins according to their privileges. While ordinary users can only perform tasks related to their computing sessions, domain admins have the authority to perform tasks within a domain, and superadmins perform almost all tasks throughout the system. A user belongs to one domain and can belong to multiple projects within a domain.

  • Compute session, container: An isolated virtual environment in which your code runs. It looks like a real Linux server with full user rights, and you can’t see other user’s session even if it’s running on the same server as your session. Backend.AI implements this virtual environment through a technology called containers. You can only create compute sessions within the domain and projects to which you belong.

  • Domain: This is the top layer for authority and resource control supported by Backend.AI. For companies or organizations, you can view domains as an affiliate and set up per-domain (or per-affiliate) permissions and resource policies. Users should belong to only one domain, and can create sessions or do some other jobs only in their own domain. A domain can have one domain admin or more, who can set policies within the domain or manage sessions. For example, if you set the total amount of resources available in a domain, the resources of all containers created by users in the domain cannot be greater than the amount set.

  • Projects: A hierarchy belonging to a domain. Multiple projects can exist in one domain. You can think of a project as a project working unit. A user can belong to multiple projects at the same time within a domain. Compute sessions must belong to one project, and users can only create sessions within their own projects. Domain admins can set policies or manage sessions for projects within the domain. For example, if you set the total amount of resources available within a project, the resources in all containers created by users in the project cannot be greater than the amount set.

  • Image: Each container has a pre-installed language-specific runtime and various computational frameworks. The state of such snapshots before they are executed is called an image. You can choose to run an image provided by the cluster admin or create your own image with the software you want to use and ask the admin to register it.

  • Virtual Folder (vfolder): A “cloud” folder that is always accessible and mountable in a container on a per-user basis, regardless of which node the container runs on. After creating your own virtual folder, you can upload your own program code, data, etc. in advance and mount the folder when you run the compute session to read from and write to it as if it is on your local disk.

  • Application service, service port: A feature that allows you to access various user applications (eg DIGITS, Jupyter Notebook, shell terminal, TensorBoard, etc.) running within the compute session. You do not need to know the container’s address and port number directly, but you can use the provided CLI client or GUI Web-UI to directly access the desired daemon of the session.

  • Web-UI: A GUI client that is served as a web or stand-alone app. You can use the service after logging in by specifying the address of the Backend.AI server and entering the user account information.

  • Local wsproxy: Proxy server built into the Web-UI app. Local wsproxy converts general HTTP requests between the server and Web-UI app to websocket and delivers the messages. If the Web-UI app loses its connection to wsproxy or the wsproxy server is dead, it will not be possible to access services such as Jupyter Notebook and Terminal.

  • Web wsproxy: In the case of the Web-UI provided in a web, the built-in server cannot be used due to the nature of the browser. In this case, you can use services such as Jupyter Notebook, Terminal, etc. in the web environment by making the wsproxy server as a separate web server so that the Web-UI app can see the web wsproxy.

Backend.AI feature details



NVIDIA support

Container-level multi GPU

(Enterprise) Fractional GPU sharing & scaling

Multiple CUDA library version support (8.0 to 12.2)

GPUDirect storage support

NGC (NVIDIA GPU Cloud) image integration

GPU-GPU Network autoconfiguration

NPU Support

* Graphcore IPU
* Rebellions ATOM
* FuriosaAI Warboy
* Hyperaccel LPU


On-premise installation on both bare-metal / VM

Hybrid cloud (on-premise + cloud)

Polycloud (multi-cloud federation)


Unified scheduling & monitoring with GUI admin

Per-user (keypair) resource policy

(Enterprise) Per-project resource policy

Availability-slot based scheduling

(Enterprise) Utilization based resource management

Cluster partitioning

Resource groups by H/W spec and usage

(Enterprise) Access control of users to
resource group

(Enterprise) Access control of project to
resource group


Sandboxing via hypervisor/container

Access logs for each user

Per session (container) logs


GUI web interface

(Enterprise) Admin GUI web interface

(Enterprise) MLOps pipeline web interface

Data management

EFS, NFS, SMB and distributed file system
(CephFS, GlusterFS, HDFS, etc)

(Enterprise) Storage solution integration:

* PureStorage FlashBlade
* NetApp ONTAP & NetApp XCP
* Weka.io
* Dell EMC
* IBM SpectrumScale (GPFS)
* Vast Data
* KT Cloud NAS

Fine-grained Access control to data by user/project

Per user/project based storage quota managment*

Developer support

Universal programming languages (Python, C/C++, etc)

Interactive web apps (Terminal, Jupyter, VSCode,
MLFlow, Microsoft NNI, R Studio, etc)

Offering various type of computing sessions tailored
to specific purposes
* Interactive
* Batch
* Inference
For data scientists

Concurrent user of multiple versions of libraries:
Tensorflow, PyTorch, etc

Concurrent user of multiple versions of libraries

Periodic update of ML libraries

Customer support
On-site installation (bare-metal / VM)

Configuration support (on-premise + cloud)

Admin/user training

Support for updating to latest version

Priority development and escalation

Customized container image / kernel or kernel
  • ONLY Available on storage supports directory quota management

Accessible menu by user role


  • Pages with * mark are in Administration menu.

  • Features only for admin are listed in admin menu.

page \ role












Import & run



Data & Storage



Agent Summary