The AI Systems, designed to support the scientific community with a focus on AI research, offers various computing and storage components for AI, data analysis, and big data research and application. Multiple generations and architectures of components are made accessible and usable in an integrated and coherent way to meet different user needs and expectations in a unified way. Currently, they are not or only partially suitable for inference applications.
For an overview of currently available compute resources, including GPU accelerators, see Performance Scope. Integration with the LRZ Data Science Storage (DSS) infrastructure makes it possible to store, retrieve and share large amounts of data on storage systems that are accessible across the entire LRZ system landscape. Dedicated storage systems for AI and machine learning meet the requirements for high performance and low latency for AI applications (see scope of services for details).
Access to AI Systems is not only made possible via traditional SSH-based login nodes (along with HPC-standard batch job workload management), but also via a user-friendly web interface. In addition to job and file management in the web browser, the latter also allows the creation of graphical user sessions and thus offers popular development environments for interactive work. All execution environments are built on container virtualization, enabling highly adaptable, yet consistent and repeatable workflow design.
Scope of Services and Service Specifics
The AI Systems consist of hardware and software components of varying performance. The full current configuration can be found here: https://doku.lrz.de/2-compute-10746641.html
- The latest compute components include:
- NVIDIA HGX H100: 30 nodes with 96 usable CPUs each, 768 GB of system memory, 4xNVIDIA H100 GPUs (94 GB of GPU memory)
- NVIDIA HGX A100: 5 nodes with 96 usable CPUs each, 1 TB of system memory, 4xNVIDIA A100 GPUs (80 GB of GPU memory)
- NVIDIA DGX A100: 4/1 node with 252 usable CPUs each, 2/1 TB system memory, 8xNVIDIA A100 GPUs (80/40 GB GPU memory)
- Storage systems: high-performance SSD-based data science storage (DSS) systems (IBM ESS3200) for AI/machine learning workloads; up to 4 TB per eligible LRZ project (see section 6.1.3); Access to all other LRZ-DSS containers/systems is also possible, but separate costs may be incurred here.
- User Frontends:
- SSH login nodes with direct command-line access to, among other things, the Enroot container runtime (container virtualization management), Slurm Workload Management (for batch job submission and management), and DSS storage management tools
- Web-based login nodes with job and file management utilities that are available directly in the web browser and also allow the establishment of graphical user sessions that development environments such as Jupyter Notebooks, JupyterLab or RStudio Server offer for interactive work
For further information, please refer to the entry page of the service documentation: https://doku.lrz.de/ai-systems-11484278.html
Requirements
A valid LRZ user account is required to access the AI Systems.
Potential users of AI Systems must demonstrate the research background/focus of their intended work or use case. This can be done by linking the user account to an AI-eligible LRZ project, i.e. an LRZ project that has the authorization to manage users of the AI Systems. Master users of such LRZ projects can manage permissions of individual user accounts by adding or removing them from the corresponding AI Systems groups.
Alternatively, access can be requested via the BayernKI initiative: https://www.ki-in-bayern.de/
Usage guidelines
The service-specific usage guidelines must be observed: https://doku.lrz.de/ai-systems-11484278.html#regulations
In addition, the usage guidelines of the following services must be observed:
- MWN connection with VPN to MWN service option
- Data Science Storage (DSS)