The most arithmetic ever sustained on scattered data

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]

22.0 petaflops

Held by LineShine. Set 2026-06-22, and has stood for 127 days. Written down here 2026-10-09.

The fastest computer on earth, elsewhere on this page, is doing 2,198 petaflops. The same machine, on the same day, doing this instead: 22. One per cent. Neither number is a lie, and the gap between them is the most useful thing on this page. The headline benchmark solves a dense system of linear equations - every number it needs is sitting next to the last one, the machine fetches memory in long tidy runs, and the arithmetic units stay fed. This one solves a sparse system, which is what a simulation of anything physical actually produces: a grid where each cell is coupled to its twenty-six neighbours and to nothing else, so the numbers a processor needs next are scattered across memory it has to go and find. The arithmetic is then almost free and the fetching is the whole job. A machine built to do arithmetic rather than to move data spends ninety-nine per cent of itself waiting, and that is not a flaw in the benchmark - it is the reason a weather model, a crash simulation or a reactor code does not get anywhere near the number on the press release. The part worth staring at is what happens to the ranking when the question changes. Supercomputer Fugaku, built in Japan and six years old, is ninth on the fastest-computer list and third in the world at this, because it was designed around getting data to its processors rather than around how many of them there are: it returns 3.6 per cent of its own headline speed where the new leader returns 1.0. Nothing about which computer is fastest survives that intact. The honest answer is always "at what", and the same body publishes both answers, twice a year, on the same afternoon.

That description was written when this frontier was opened, on 2026-10-09, and is never rewritten. Where it and the number above it disagree, the number is the reading and the description is history.

What counts

The HPCG figure of the number one system on the current HPCG list, published by TOP500 alongside the TOP500 list itself - sustained arithmetic on the High Performance Conjugate Gradient benchmark, whose data sits scattered through memory rather than lined up in a dense block. The list states the figure in teraflops and the publisher's own release prose states it in petaflops; this frontier carries petaflops, so that it can be read against the fastest-computer record here without arithmetic. Four rules belong on the number rather than under it. First, this is not the fastest-computer record re-pointed: it is a different benchmark run on the same machines, it is published as its own ranked list, and the two orderings genuinely differ - Supercomputer Fugaku is third in the world at this and ninth on the TOP500 list of the same edition - so a change in one is not a change in the other. Second, the edition is found from the evergreen index and never from a dated address, which is the trap the fastest-computer record carries a warning about: top500.org/lists/hpcg/ links every edition and the newest one it links is the list to read. Third, the publisher's rank decides. The edition page publishes the top ten rather than the whole list, so a maximum cannot be taken over the column from here and is not taken: the reading is the rank one row, cross-checked against the sentence the release prose prints naming the leader and the figure, and the ten rows are sorted in a script only to confirm the ordering is real and the column is not misaligned. That is the Prime Pages rule of 2026-10-01 - where the publisher has already ranked the data, read the rank. Fourth, a machine must have submitted a result to appear. HPCG is optional, and the June 2026 release says in as many words that the JUPITER Booster system has not submitted one, so this is the most anyone has submitted and had published rather than the most any machine could do. It is also the current list rather than an all-time archive: a leader dropping off the list can lower this number, and that is the publisher's answer rather than an error.

What moves it

TOP500 publishes an HPCG list whose leader carries a larger figure. The list arrives on a calendar rather than as submissions clear - every June and every November, with the TOP500 list - so this number can only move twice a year, and the next edition is due around November 2026.

Where to check it: https://www.top500.org/lists/hpcg/

The program reads it from https://www.top500.org/lists/hpcg/2026/06/, which is the same figure in a form a script can parse.

Last checked: 2026-10-26. The source itself last measured this: 2026-06.

Previous: Theorems a computer has checked   Next: The fewest mistakes a machine makes transcribing a meeting

[ Front page ] [ The edge ] [ Finished ] [ About ] [ Links ]