Abstract:Supercomputing power has now become an important index of science and technology competitiveness of a country, and also a great support to the economic and social development and national security. Moreover, with the drastic development of Internet, big data, cloud computing, and virtual reality technologies, supercomputing will bring about profound changes to production and life styles of modern societies. Nevertheless, contemporary supercomputing is facing grand challenges in terms of low efficiency, high power consumption, poor robustness, and difficulties in applications. General-purpose supercomputing based on commercial processors is not likely to solve the problems completely. Based on our explorations in multiscale simulation of complex systems, we suggest that keeping the structural and logical consistency between the computed system, model, software, and hardware is an effective way towards high efficiency and scalability of supercomputing in the future. In fact, multiscale structure is ubiquitous in natural and industrial systems, which is, after all, a reflection of the hierarchical nature of the physical world and is, therefore, of general significance to a wider range of applications. For classical mechanical and thermal processes, the computational complexity usually decreases with decreasing scales, while the computational cost increases exponentially. Fortunately, with reasonable coarse-graining based on meso-scale models, this cost can be reduced by several orders by developing additive discrete methods which can achieve good accuracy, efficiency, and scalability for supercomputing. On the other hand, computation at large scales can be accelerated by introducing stability constraints to the dynamics models. Multiscale computer hardware conforming to this software framework is expected to have high efficiency and speeds up the computation even further. We have already explored the feasibility of this multi-scale strategy for model, software, and hardware developments in different applications, and have achieved very encouraging results. We suggest that this strategy should be given enough considerations in the development of future supercomputers.