Computational physics, as a discipline of computational science, is essential for modern scientific research. It is a discipline in which supercomputers are used to perform computation based on the known laws of physics, and simulate conditions which can hardly be reached with current technology, or realized by experiments, or where massive experimental data are to be processed.
Computations in different research fields have different demands for computational resources, but generally applications for computational physics require extremely high floating point execution power and memory bandwidth, therefore such computations can be only accomplished in relatively large clusters.
This presents further high requirements for the computational network bandwidth and latency.
Computational solid state physics (Material science)
Quantum Espresso, Material studio, SIESTA, CASTEP, DMol3, phonon, TB-LMTO
FDTD Solutions, Acceleware, Meep
Computational particle physics
CompHEP, FeynArts, LoopTools
Application Performance Features
With high compute accuracy and efficiency, VASP software is based on the first-principle theory and performs computation using the pseudopotential plane wave method, and is widely used in computational for solid state physics, materials science, particle physics, etc. It is also used even in astrophysics, geophysics, etc., and so it is one of the most important scientific applications. Inspur have theoretically and algorithmically conducted intensive studies on VASP for many years and gained deep insights.
The run time features of VASP are described as follows by taking magnetic materials - a hot research topic currently - as an example.
Number of ions
Number of plane wave basis sets
Number of bands
Number of K points
After optimization: on CPU clusters VASP can be extended to ~600 cores with a single K point.
After optimization: the efficiency of the GPU cluster is much higher than the CPU cluster
While running, VASP can reach a floating point execution speed of 400GFlops in a single node, a memory bandwidth of 100GB/s and an inter-node network bandwidth of 1GB/s. Therefore, it is a floating point intensive, memory bandwidth sensitive and network bandwidth intensive application. A computer network with high floating point execution capacity, high memory frequency and extremely low latency can help deliver VASP’s computational potential. After being optimized, VASP can reach a good acceleration ratio and it is very suitable for accelerating computation by using GPU cards.