While this is the #1 supercomputer in the world, its conjugate gradient performance is not the best. This is interesting because it was said that for scientific computing, iterative solving algorithms are the most important uses for supercomputers.
Metalbird
From what I remember from listening to a talk by Dongarra was that some computers have been designed to do exceptionally well on the LINPACK benchmark, while the conjugate gradient test is more relevant of high performance scientific computing. Conjugate gradient is iterative, uses sparse operations, and requires more memory accesses than LINPACK does. So computers can be designed to perform well on LINPACK, and do well on some types of scientific computing, dense calculations, but may be less efficient at other types.
While this is the #1 supercomputer in the world, its conjugate gradient performance is not the best. This is interesting because it was said that for scientific computing, iterative solving algorithms are the most important uses for supercomputers.
From what I remember from listening to a talk by Dongarra was that some computers have been designed to do exceptionally well on the LINPACK benchmark, while the conjugate gradient test is more relevant of high performance scientific computing. Conjugate gradient is iterative, uses sparse operations, and requires more memory accesses than LINPACK does. So computers can be designed to perform well on LINPACK, and do well on some types of scientific computing, dense calculations, but may be less efficient at other types.