What do supercomputers do every day? They do experiments that are impossible in the real world. For example, they help predict the weather and find new medicines.

In 2020, Japan's supercomputer Fugaku showed how drops from coughs spread. Fugaku was the fastest computer in the world four times in a row. It can do about 440 quadrillion calculations in one second!

Supercomputers do experiments that are impossible in real life. We cannot put the Earth's atmosphere in a laboratory or test millions of drug candidates one by one, but we can do both inside a computer. Japan's Fugaku, developed by RIKEN and Fujitsu, was ranked first in the TOP500 list four times in a row, from June 2020 to November 2021. It could perform about 440 quadrillion calculations per second.

Supercomputers divide the air into small boxes to forecast the weather, simulate earthquakes and tsunamis, and narrow down candidates for new medicines. In 2020, Fugaku simulated how droplets from coughing and talking spread, and these images appeared widely in the news. However, being first in the ranking does not mean being best at every task.

Nobody can put the planet's atmosphere in a laboratory or synthesize millions of drug candidates. A supercomputer can simulate both, and running experiments that would be impossible in reality is its main job. Speed is measured in calculations per second. The TOP500 list, published twice a year, ranks machines using the LINPACK benchmark.

Japan's Fugaku, developed by RIKEN and Fujitsu, topped the list four times in a row from June 2020 to November 2021, at about 440 quadrillion calculations per second. Since 2022, exascale machines that exceed a quintillion calculations per second, mostly in the US, have held the top places. Supercomputers divide the atmosphere into grid cells to predict weather, model how earthquake waves and tsunamis spread, screen possible medicines and, more and more, train AI.

In 2020, Fugaku was put to work early to simulate how droplets from coughing and talking spread, and the resulting images were widely broadcast. Those simulations rest on assumptions about evaporation and ventilation, so they are better for comparing risks than for giving exact infection probabilities. LINPACK, for its part, measures only one aspect of performance. Memory speed, communication and energy efficiency also matter, so the "best" machine depends on the task.