Large supercomputers need large amounts of storage, which means the network needs to not only connect the individual computers together, but also provide fast access to the data. These days, much of the computing work is done by GPUs or other accelerators as we call them. For many of the kinds of math problems we deal with in science, these accelerators can solve them much faster than the regular processor or CPU. Of course, the companies manufacturing computer chips are now adding GPUs and accelerators to the CPUs, so we don’t have to buy separate parts. All of these parts have different tradeoffs, like cost, performance, and how much heat they generate, so we pick a set of parts that complete a particular workload quickly for a reasonable amount of money. (Workloads are the mix of different jobs a supercomputer does, such as predicting the weather or sorting through images.)
The supercomputer is only part of the picture. It’s where a lot of computing work gets done, so it’s an important resource. There’s also the staff that maintain the hardware, support the users, and run the data center where the supercomputer lives. The software used by the researchers also needs to be developed and frequently updated to take advantage of new technologies; for example, cool new things like CPUs that have built-in accelerators. Researchers eventually need to publish and present their results, so we provide ways for them to get their data from the supercomputer to their laptops quickly so they can do data analysis and write their papers.
Everything I’ve described are things that are involved with supercomputers and how we make them useful for science.
