Guide
How supercomputers work
Nine chapters on what actually separates a supercomputer from a fast computer. It starts from one fact (these machines are built from parallelism rather than speed), and follows the consequences through the node, the network, the file system, the cooling plant, the programming model and the operations team. Written to be read in order, and to stand alone if you arrive in the middle. About 116 minutes end to end.
Every chapter links into the systems, parts and companies in this dataset, so a claim about how something works can be checked against machines that actually did it. Where a chapter states a number, the system page behind it carries the source.
- 01 What actually makes a supercomputer different A supercomputer is not a fast computer. It is a very large number of ordinary computers wired together tightly enough to work on one problem at once, and almost every design decision in the machine follows from that one fact.
- 02 Inside a node A supercomputer node is built around one problem (arithmetic units are cheap and data movement is expensive), so almost every feature of a modern node is a scheme for keeping the arithmetic fed.
- 03 The interconnect The network between nodes is the part of a supercomputer that most determines whether it works, and the part that most distinguishes it from a warehouse full of servers: it is judged on latency and on worst-case behaviour, not on bandwidth.
- 04 How supercomputer performance is measured The number these machines are ranked on measures dense matrix factorisation, which almost nobody runs, and the gap between that number and real application performance is often a factor of fifty.
- 05 Storage and I/O Compute has grown far faster than the storage under it, so the practical limit on many large runs is not how fast the machine calculates but how fast it can write down what it calculated.
- 06 Power and cooling A leadership-class system draws 20 to 40 megawatts, which is a small town, and the constraint on the next generation is now more often the substation and the heat rejection plant than the silicon.
- 07 Programming a supercomputer There is no compiler flag for this. Code must be written to express parallelism at three different levels at once, and porting a large scientific application to a new accelerator is measured in years of person-effort.
- 08 Operating a supercomputer At nine million cores something is always broken, so a large facility is run less like a computer and more like an industrial plant, with scheduling, allocation policy, failure budgets and a lifecycle measured in years.
- 09 AI clusters versus HPC systems They look like the same machine and are not: different arithmetic, different failure tolerance, different communication pattern, and a disclosure culture so different that most of the largest compute built in the last three years appears on no public ranking at all.