GPU HPC and AI Hackathon 2026 at IDRIS in May 2026
This page was translated by an AI (LLM) with a cursory human check and is awaiting full review.
For the sixth consecutive year, IDRIS is organising, in conjunction with NVIDIA, a GPU HPC and AI Hackathon.
It will take place on May 12th (remotely) and from May 19th to 21st 2026 (at the IDRIS premises). The registration deadline is set for March 10th.
GPU Hackathons allow teams of three or four developers or researchers to improve the performance of their HPC or AI codes under the guidance of one or more mentors, who are experts in GPU programming and come from universities, national laboratories, computing centres, governmental institutions, and manufacturers.
It is not necessary to be a user of IDRIS to participate in this event, so feel free to share this message with your colleagues. You will have the opportunity to access IDRIS's computing resources to test your code on recent GPUs, including NVIDIA H100s on Jean Zay and NVIDIA GB200 NVL72s on the DALIA prototype machine.
For HPC projects, it could be a matter of starting to port a code to GPU using OpenACC, OpenMP, Kokkos, GPU libraries, CUDA, etc., or optimising an existing GPU implementation. For large codebases, it is recommended to prepare a representative mini-application to be more efficient during the event.
For AI projects, this Hackathon offers a unique opportunity to explore fine-tuning, developing, or evaluating state-of-the-art models such as LLMs and GenAI applications. You can optimise your models using advanced techniques such as mixed precision, multi-GPU or multi-node parallelisation (DDP, PP, TP, CP), using basic frameworks such as PyTorch or TensorFlow, or specialised tools such as DeepSpeed, Megatron-LM, Nanotron, or NVIDIA NeMo. Although full pre-training is out of reach within the Hackathon's timeframe, preparing a pre-training code for large-scale models is still possible.
Registration is done via the event page (in English): https://www.openhackathons.org/s/siteevent/a0CUP00003Ahxc22AB/se000449.