White paper : Enabling Scalable CFD
Simulations in the Cloud

Born from a collaboration between EDF R&D MFEE, NVIDIA, AWS, and ANEO, this white paper shows how to port and optimize code_saturne on AWS GPU instances—from standing up a cloud HPC stack (ParallelCluster, Terraform, Amazon DCV) to fine-grained kernel tuning with Nsight Systems & Nsight Compute. You’ll get concrete takeaways on gains, trade-offs, and best practices, plus benchmarks on L40S (G6e) and H100 (P5/P5en).

On the agenda

• Why & how move to GPUs on AWS? Motivations, target architecture, and tooling to run code_saturne reliably and reproducibly in the cloud.  
• Architecture & deployment: AWS ParallelCluster v3.12 (Slurm, EFA), a GPU dev workstation via Amazon DCV, IaC with Terraform, and an access portal (Admin / End-User profiles).
• GPU optimization methodology: Top-down profiling (Nsight Systems) and bottom-up (Nsight Compute), memory management (Unified Memory, prefetch), keeping data on GPU, reducing allocations.
• Results, benchmarks & best practices: C016/F128 cases, L40S vs H100 comparisons, MPI/halo-exchange bottlenecks, acceleration levers (hot kernels, single/mixed precision where valid), and a next-steps roadmap.