While the kernel of the HPCG, based on the Conjugate Gradient Method with a symmetric Gauss-Seidel preconditioning, has obviously very low arithmetic intensity, that can be assessed using analytical methods, programs for Numerical Weather Prediction and atmospheric physics seem to be the most complex of standard simulation, both because the physical environment is a complex mixture of gases, particles, and water as ice, snow, rain, fog, vapor, and the number of observed parameters is also much larger than in other scientific and engineering simulations. Unlike the HPCG, based on pure linear algebra, weather prediction programs use complex mathematical functions (exp, log, x^z, ...) making the computation highly non-trivial.
Moreover, there is 10 supercomputers in Top100 (June 2019) with total more than 1 million of cores and more than 50 PFlop/s, making the weather segment of HPC one of the most important fields.
This is why the initial study of LOWAIN feasibility concentrated to the NWP. As the program to start, we have selected the WRF (Weather Research & Forecast) program, perhaps the most important open program in the field (ConUS - WRF analysis of the Continental U.S. domain, Oct. 24, 2001 - is one of the main HPC benchmarks).
The WRF is a huge program, and hence the research was limited to 3 most time consuming and complex modules of the WRF:
The arithmetic intensity, defined as the the ratio of the number of performed operations and the number of bytes that crossed the processor-memory interface was measured as the function of the cache size on a virtual computer with a variable cache size (the large is the cache, the smaller number of data need to cross the processor-memory interface).
A preliminary detailed report can be found here. The research has shown that the arithmetic intensity of the WRF is not much larger than that of the HPCG, which suggests that LOWAIN might be an advantageous architecture for the NWP.
The results were presented at the General Assembly of the European Geoscience Union in Wien, March 2019 (download).