Modeling the Climate
A low-power supercomputer could lead to ultrahigh-resolution climate models
Source: “Towards Ultra-High Resolution Models of Climate
John Shalf et al.
International Journal of High Performance Computing Applications 22: 149-165
Results: Engineers and scientists at Lawrence Berkeley National Laboratory have designed a low-power supercomputer that can resolve climate models to the kilometer.
Why it matters: Today’s supercomputers aren’t nearly powerful enough to run algorithms that predict future weather with kilometer-scale resolution–algorithms that would allow researchers to improve the overall accuracy of climate models and better inform local decisions about how to adapt to global warming. The proposed supercomputer would be not only more powerful but hundreds of times more power efficient than any other supercomputer, making such calculations cost effective.
Methods: The researchers used chip design software made by Tensilica, a chip manufacturer in Santa Clara, CA, to build a processor with only the functions necessary for weather modeling, as opposed to the general-purpose chips commonly used in today’s supercomputers. They are also custom-designing the chip’s memory and the circuitry that connects its 32 processors–or “cores”–to reduce inefficiencies and minimize power consumption. The climate model that will run on the machine splits the globe into 20 million cells–one for each core in the supercomputer–and will be able to simulate the movement of storm systems and weather fronts.
Next steps: The researchers still need to finalize the design of the 20-million-core supercomputer. They also plan to run the climate model on simulations of the supercomputer, looking for problems with either the hardware or the software and opportunities to optimize performance and energy use.