This subsection shows that simulation programs in CFD, NWP, and other fields use typically just a very small fraction of the computing power of the system.
The previous subsection shows that the main reason why supercomputer simulations are computationally so inefficient is insufficient memory bandwidth of processors and accelerators. The data presented in this subsection suggest that the use of the memory bandwidth is not as high as it should be, thus giving space for improvements.
Modern processors and accelerators achieve rather large memory bandwidth only when working with relatively small high bandwidth memories (HBM) tightly coupled with the processor, usually in the same package. If larger external memory is needed, the memory bandwidth drops to a fraction of the HBM value.