Cluster Architecture#

A high-level look at how the LTCI cluster is built and what it can do. Exact per-node specifications (hostnames, core counts, memory) are in the Infrastructure documentation, available to cluster users.

At a glance#

System overview#

The cluster follows the classic high-performance-computing model:

• Login nodes: where you connect, prepare your code, and submit jobs.
• Compute nodes: where jobs run, grouped into partitions by hardware.
• Shared storage: one filesystem visible from every node.
• A fast network that ties it together and provides internet access.

Compute partitions#

Compute nodes are grouped into partitions by GPU generation. You choose a partition when you submit a job, and the scheduler (Slurm) runs your work on a free machine in it.

Several dedicated partitions are reserved for specific research teams and funded projects.

Each compute node pairs several GPUs with multi-core CPUs and a large amount of memory. Generations span NVIDIA Pascal (P100) to Hopper (H100), so the cluster handles both everyday development and demanding, latest-generation work.

Software & environment#

The cluster runs modern machine-learning and scientific workloads out of the box:

• Frameworks: PyTorch, TensorFlow, and JAX, installed in your own environments (venv, uv, or conda) or pulled as containers.
• GPU stack: multiple CUDA toolkit versions and cuDNN, managed through environment modules (Lmod).
• Containers: Apptainer runs any Docker or NVIDIA NGC image with your normal user permissions, so you can bring a fully reproducible stack.
• Interactive work: JupyterHub gives web notebooks directly on cluster GPUs, alongside interactive shells and VS Code Remote-SSH.

See the Software documentation for the full catalogue and guides.

Scaling & network#

• Single-node multi-GPU training is straightforward, since most nodes hold several GPUs.
• Multi-node distributed training is supported. Nodes are connected over Ethernet, and the newest generations (H100, A100, L40S) add RDMA over Ethernet (RoCE) for lower-latency scaling.
• Full internet access from compute nodes, so you can download datasets, pull container images, and reach external services such as Hugging Face Hub and Weights & Biases directly from your jobs.
• The cluster is reachable only from the Télécom Paris network, on campus or over the VPN.

Storage#

All nodes share the same NFS filesystems, so your files appear on every node, login or compute, with nothing to copy.

See the Storage documentation for quotas, backups, and usage policies.