The most questions ever answered against a hundred terabytes of data

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

653,611,141 queries an hour

Held by Tencent Cloud Data Lake Compute. Set 2026-04-18, and has stood for 192 days. Written down here 2026-10-04.

Every other speed on this page lets the entrant choose how big the problem is. The fastest computer picks the size of the matrix it inverts, the graph search picks the size of the graph, and the server doing ordinary work picks how many copies of each program to run - which is why those numbers all carry a caution about comparing one row with the next. This one does not need the caution, because the size is fixed before anyone starts and it is written into the name of the measurement: a hundred terabytes of data, the same hundred terabytes for everyone, and the only question is how fast a system can interrogate it. The interrogation is ninety-nine queries over a model of a retailer that sells through shops, a catalogue and a website - the kind of question a company actually asks its own records, which is to say how much of what sold where to whom and against what the year before. The record holder is a hundred and forty-two rented machines with four and a half thousand processor cores and fifty-four terabytes of memory between them. It loaded the hundred terabytes in seven and a half minutes, answered all ninety-nine questions once in under three minutes, then answered them four at a time, twice, in seven and a half minutes each. One honest thing about the figure, which the council's own method makes unavoidable: it is a composite and not a count. The score is the size of the database multiplied by the number of queries in a throughput run and divided by the average of four timings - loading, one stream, the concurrent streams, and the updates that run between them - so a figure of six hundred and fifty million queries an hour belongs to a machine that in fact answered eight hundred and ninety-one queries in about twenty minutes. The hundred-thousand multiplier is the database measured in gigabytes. That makes the number meaningless on its own and meaningful against another result at the same size, which is exactly what the council says in the note it prints above the list, and it is a more honest arrangement than most of the scores on this page manage. The rest of the method is the same one that produced the transaction record elsewhere here: an industry body whose members are the vendors being ranked, an independent auditing firm signing the result, and the price of the system published beside its score - seven and a quarter million yuan, for this one.

That description was written when this frontier was opened, on 2026-10-04, 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 largest figure in the Performance (QphDS) column of the hundred-terabyte group on the Transaction Processing Performance Council's list of all published TPC-DS results - the council's composite decision-support score, which it writes QphDS@100000GB. Four rules belong on the number rather than under it and all four are the publisher's own. Major versions of TPC-DS are not comparable, which the council states at the top of the list, so this frontier counts the version the list serves today - Version 4 - and the reading begins by confirming that the page still says V4 in its own heading; if it ever says V5 this frontier needs a decision before it needs a reading, and the successor is a new frontier rather than a re-pointing of this one. Results measured against different database sizes are not comparable either, which is why the council groups them and why this frontier names one group: a hundred terabytes, the largest it publishes and today the only one Version 4 holds. A result the council marks Withdrawn does not count, while one marked Result In Review does - the council ranks those in its own list, as it does on SPC-1, so this follows the publisher rather than substituting a stricter standard of its own, and the status is read from the winning row's detail page. And the group is ordered by performance, but the reading is taken by sorting the column in a script and never by trusting the order.

What moves it

The council publishes a result in the hundred-terabyte group carrying a larger figure. Submissions arrive as they clear the council's process rather than to a conference calendar, so this can move in any week - and the group is young enough that all three results it holds arrived inside the last thirteen months.

Where to check it: https://www.tpc.org/tpcds/results/tpcds_results5.asp

Last checked: 2026-10-26. The source itself last measured this: 2026-08-19.

Previous: The most ordinary work one computer can do   Next: The least time ever taken to train a four-hundred-billion-parameter model

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