Simulation

Hatem Ltaief : Principal Research Scientist, KAUST
Qinglei Cao : Assistant Professor, Saint Louis University
Slides
Geostatistical and spactal simulation (not less than 100 km, and daily only with asumtion of axial symetry)
Matern Covariance Function
Statistical Emulator (approximate complex climate simulation)
Cholesky Factorization, A=L L^t (most of the computing time)
Compute determinent of the matrix (dense matrix but with data sparsity structure)
They can use low rank approximation
On Summit, ALPS and LEONARDO
Tile scheduling of the Cholesky Factorization
Portability (Intel, ARM, NVidia, AMD, CPU, GPU) (because they rely on vendors BLAS libraries)
New steps since the presentation of 2022
Do not oversolve or overstore
cuTile and cuDTX presented by Stephen Jones will be integrated
Tasks based runtime (they are based on PaRSEC which is 15 years old)
Many matrices arising in applications have blocks of relatively small norm and can be replaced with reduced precision
As always data movement is really expensive
Speed up of 3.4 on LEONARDO (A100)
Speed up of 5.2 on ALPS (GH200) 0.739 EFlop/s
No AI comparison, problem with overfitting, expensive to retrain
But, they want to apply mixed precision algorithms to AI models
Using FP64 Emulation is possible to integrate the remaining FP64 blocks