Lyra: EPFL's new GPU cluster now open for beta testing
We are pleased to announce that, following several weeks of testing and validation, Lyra, EPFL's new GPU cluster, is now open for beta testing to the entire HPC/AI community.
The deployment of Lyra has also been an opportunity for remarkable collaboration across EPFL. The SCITAS team has worked closely with colleagues from VPO responsible for the datacenter infrastructure, who have been involved since the design phase, and have helped to address the requirements of this new generation of high-density computing systems. We are particularly proud that this collaboration has resulted in the installation of the first Direct Liquid Cooling (DLC) rack in the CCT datacenter, paving the way for increasingly dense and energy-efficient computing infrastructure at EPFL.
We have also greatly benefited from the collaboration with our colleagues from EcoCloud, who actively participated in the cluster acceptance phase, helping us validate its energy performance and assess the environmental impact of this new infrastructure.
Lyra can be accessed via SSH through the following frontend:
lyra.hpc.epfl.ch
Currently, the cluster provides two GPU partitions:
- The rtx6000 partition consists of 14 compute nodes, each equipped with 8 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with 96 GB of GPU memory each, 2 AMD EPYC 9535 64-Core processors for a total of 128 CPU cores, and 768 GB of system memory. This configuration corresponds to 16 CPU cores per GPU and approximately 5.7 GB of system memory per CPU core.
- The b200 partition consists of 8 compute nodes, each equipped with 8 NVIDIA B200 GPUs, 2 Intel Xeon Platinum 8568Y+ 48-Core processors for a total of 96 CPU cores, and 1536 GB of system memory. This configuration corresponds to 12 CPU cores per GPU and approximately 15.5 GB of system memory per CPU core.
Please take these CPU and memory ratios into account when sizing your resource requests. It is mandatory to specify the appropriate partition when submitting jobs. For example:
More detailed documentation will be available in the coming days, as part of the user portal, and will be linked from the login message.
During the beta-testing phase, all usage on Lyra will be free of charge until 30 September 2026. No invoices will be issued for usage during this period. Tentative prices will be displayed in the login message, and a separate communication will follow with the official costs as soon as they are confirmed by the financial services. Please keep in mind that Lyra is still undergoing its final acceptance. The service may therefore be subject to configuration changes, interruptions, or short-notice maintenance operations. Running jobs may be terminated and unsaved data could be lost, so we recommend that workflows save checkpoints regularly. We want this beta phase to be accessible to the broader EPFL scientific computing community. Users who are not yet familiar with the Slurm scheduler or the scientific software environment available on our HPC/AI systems can benefit from dedicated support to port and test their scientific workflows on these latest-generation GPUs.
Your feedback during this phase is particularly valuable. Please report any issues related to system stability, job execution and scheduling, GPU performance, the software stack, or unexpected behaviour and performance regressions to 1234@epfl.ch, including the keyword "HPC" in the subject line.
We look forward to seeing your scientific workloads running on Lyra and receiving your feedback, as we take this step on building the future Integrated Computing Platform.