12.4.6 : Simulation
- Tuesday, (4:30 PM - 4:55 PM CET) A New Era of AI-Driven Electronic Design Automation on Accelerated Computing [S62846]
- Paul Cunningham : Senior Vice President and General Manager, Cadence Design Systems, Inc.
- Slides
- Cadence : design and simulate electronic and physical system
- Computer Chips are powering everything
- Generative AI based execution and optimization
- Simulte chips befare they are manufactured
- Simulte fluid in data centers
- 2.8x to 6.7x on matrix inversion acceleration on GPU (Grace Hopper)
- Complexity is increasing by there are not enough engineers
- From about a day to debug a Chip problem to 10 minutes with AI assistant
- NeMo retiver and Cadence Copilot
- Without chips there is no software
- 16 months to produce a Chips and 3 months only to print the chips for real and test it
- Tuesday, (6:00 PM - 6:50 PM CET) Simulating Advanced Lunar Exploration Robotic Systems [S61635]
- Lutz Richter : Space Projects Expert, SoftServe
- Slides
- tout est dans le titre
- Wednesday, (1:00 PM - 1:25 PM CET) Accelerating Simulations of Multiscale Chemical Reactors using NVIDIA Modulus [S62060]
- Slides
- PINNS : Predict Physics
- Gaz to Liquid Chain
- Multiscale / multiphysics modelling of chemical reactors
- Iterative simulation
- SPINN : good speed up with close result to the reference
- Full scale model of the reactor
- Less than 2% error on prediciton
- Wednesday, (4:00 PM - 4:25 PM CET) Simulating Solar Eruptions on GPUs using Fortran Standard Parallelism [S61193]
- Ron Caplan : Computational Scientist, Predictive Science Inc.
- Slides
- Multiple level of physical scale and sphysical processes
- CORHEL CME : read the sun
- Magnetohydrodynamic (MHD) simulation of the sun's atmosphère
- 20 min on 4 GPU system (iterate until result corresponds to sun)
- Finite difference on a logically rectangular non-uniform spherical grid
- Fortran parallelised with MPI
- Highly memory Bound
- do concurent : say to the compiler loop may be parallelised
- Much simpler than CUDA
- Standard will stay
- Also parallelism on CPU
- No performance drop on GH200
- No du concurent reduce because gfortran was not compatible
- Less code on standard fortran
- Only a single multi GPU node for the simulation
- Next total solar eclipse : April 8 2024
- Jacobi and domain decomposition on GPU
- Wednesday, (5:00 PM - 5:25 PM CET) Spatiotemporal-Graph Neural Networks for Gravitational Wave Discovery in the Big Data Era [S62178]
- Wednesday, (5:00 PM - 5:25 PM CET) A Fully-Differentiable Lattice-Boltzmann Solver for Integrated Machine-Learning Simulation Workflows [S62237]
- Josef Winter : Researcher, Technical University of Munich
- Thomas Indinger : TUM
- Slides
- Simulation of fluid flow => use classic Lattice-Boltzmann
- But how about Hybrid approaches
- ML can gives a better first iteration for classic method
- Pytorch based
- ML Advance module updates the incoming flow
- Test with NVidia Modulus (based also on PyTorch)
- Slide 7 Backends for all aspects of the simulation
- Fourier Neural Operator (FNO) : capts the dynamic of partial differential euqlation )> train on low resulution data and use on higher reslution
- Simulate von Kàrmàn vortex Street
- Test on NVidia RTX A6000
- Slide 14 iteration of Hybrid workflow faster to converge
- Slide 15 from 407s to 204s with hybrid method
- From FNO to FNO + LBM
- Small memory overhead of the model size
- Wednesday, (11:00 PM - 11:25 PM CET) Strategies for the GPU Implementation of the OVERFLOW CFD Code [S61501]
- Chip Jackson : Research Scientist, NASA Langley Research Center
- Slides
- no record