CINECA OpenPower-based HPC Infrastructure: Some porting results
summary
Marconi100 is an OpenPower Tier‑0 cluster with ~1,000 nodes each hosting 2 Power9 CPUs (≈3 GHz), 266 GB RAM and 4 NVIDIA V100 GPUs — ~350k cores and ~4,000 GPUs total. Three production codes get ported and measured: SIM (DNS) moves to CUDA Fortran with CAF kernels and cuts time‑to‑solution to under half after layout and RDMA tuning, SeisSol rewrites kernels via YATeTo to MPI+CUDA and hits 0.5 PF (single) on 64 nodes, an LBM OpenACC run simulates 50B cells (1 TB RAM) on 500 GPUs with ~73% efficiency at 1,000 GPUs.
AI-written summary · about
chapters (7)
- 0:00Introduction to CINECA HPC
- 8:23Porting examples overview
- 17:40First code porting results
- 22:45Second code performance gains
- 26:00Third code scaling results
- 28:38Power9 performance benefits
- 29:00Q&A and closing
related talks
20mRevolution in IT
01h 19mCo-Design for High-Performance Computing Software Systems
42mSupercharge Your Performance with the Optimus Cirrus Platform
31mEnd2End: From Silicon to System Delivery
01h 50mUnlocking Modern CPU Power - Next-Gen C++ Optimization Techniques
20mHigh-Performance Computing that’s Earth Friendly
47mPrepping your Open Source Project for Multi-Platform
01h 04mThe Future of Fast Code: Giving Hardware What It Wants