How to lose a supercomputer race while standing still
The first ten places in TOP500 held steady. The entry threshold still rose. A handful of scorecards explains how both can happen.

A quiet leaderboard with a moving finish line
If you watched only the first ten places in this November's TOP500 announcement, you might have put the results down and gone back to work. Every one of those positions stayed the same. Look farther down, though, and a different story appears. A machine that finished comfortably inside the previous list can still be running just as quickly today and find itself outside the new one. How does a computer lose a race while standing still?
TOP500 announced its latest results on 16 November at SC26. LineShine led with 2.198 exaflop/s on High Performance Linpack, the benchmark used to order the list. Just 27 systems were new to this edition, a record low according to the organizers. Even so, the entry threshold rose. The results give us an unusually good excuse to look at what a ranking can tell a programmer shopping for compute time.
Source: TOP500 announcement
Race the same problem first
Imagine you and a friend have each written a program to solve a set of simultaneous equations. Before you compare times, agree on the input and on how close an answer has to be to count. Otherwise a wonderfully quick answer could simply belong to a smaller problem. Linpack gives the machines on the list a common kind of numerical work. That agreement gives the ranking meaning. It also tells you where that meaning stops.
Now imagine your own work is finding a path between two stations on a large rail network. Your program looks up a station, follows a connection, reads the next station and decides where to go next. Much of tomorrow's reading depends on what you find today. A machine can be very good at arithmetic on orderly blocks of numbers and spend much of this journey waiting to learn which piece of memory it needs. You would want to time a representative route before booking an expensive run.
What happened to number five hundred?
You can recreate the moving threshold without a supercomputer. Write five pretend scores on cards, in descending order, and call the first three your published list. Add one new score between first and second. The old third place slips off the list even though its score never changed. Now imagine reading only the champion's name. You would miss the changed entry requirement altogether. This is an illustration of how a cutoff works, rather than an explanation of any particular system's departure.
The same trap appears when you compare versions of your own software. A dashboard might tell you a service fell from the fastest ten while its response time held steady. Keep the underlying measurement next to the rank. You can then tell a colleague whether your service slowed down, other services got faster, or the group being ranked changed membership.
Write a better request for compute time
Suppose our rail network researchers get one afternoon on a large machine. They could bring a small graph whose answer they know, a realistic graph that fits in memory and a case that strains it. Recording answer quality, elapsed time and memory for all three would tell them where their program spends that afternoon. The leaderboard gives them a place to start looking. Their own timetable gives them a reason to choose.