Facts
21 numbers computed from the 1345 systems in the dataset. Each has a sentence you can paste, the method behind it, and a link to the pages it comes from. They are recomputed on every build, so they stay true as systems are added. Each also has its own page with a share card, and the whole set is at /facts.json.
Say what these are. Every figure describes the systems in this dataset, a curated sample, not the whole market. The quote sentences say "in our dataset" for that reason; please keep those words when you share. Data as of 2026-10-10.
158,158,273x
faster in 35 years
The fastest measured system in our dataset is 158,158,273 times faster than the fastest we hold from 1991: 2.20 EFlop/s for LineShine against 13.9 GFlop/s.
Largest measured Rmax among systems first operational in 1991 (Touchstone Delta (Caltech)) against the largest measured Rmax in the dataset (LineShine, 2026). Measured means a LINPACK result, not a vendor peak.
41.9M
cores in the largest system
Sunway OceanLight has 41,930,000 cores. The smallest ranked system we track, UTokyo HITAC S-3800/480, has 4.
Core counts as reported on each system row; accelerator cores are included where the source counts them, so figures are comparable within a source and not always across.
68% vs 79%
share of peak reached, GPU vs CPU-only
On the LINPACK benchmark the median accelerated system in our dataset reaches 68% of its theoretical peak. The median CPU-only system reaches 79%.
Median of Rmax divided by Rpeak, both on the same row, across 386 accelerated and 474 CPU-only systems with a measured result. A system counts as accelerated if it has an accelerator component edge.
22
systems here reached #1 on the TOP500
22 of the 1345 systems in our dataset reached number one on the TOP500 at some point: CM-5/1024, ASCI Red, Earth Simulator, BlueGene/L, Roadrunner, Jaguar and more.
Systems whose stored TOP500 pointer is rank 1. We store one pointer per system, on the edition where it ranked highest, so a system that held #1 shows once. Rank is a link-out pointer, not a copy of the list.
96% to 392%
unranked AI clusters vs every ranked system here
Clusters that appear on no ranking anywhere add up to an estimated 96% to 392% of the combined LINPACK performance of every ranked system in our dataset. It is an estimate, bracketed by measured rates on the same chips.
The anchor-bracket-v1 estimator over 177 unranked systems that disclose an accelerator count, against the summed Rmax of 848 ranked systems. No vendor peak is multiplied by an assumed efficiency.
100.00 MW
drawn by Cerebras Guyana AI Data Center
Cerebras Guyana AI Data Center draws 100.00 MW, the most of any of the 523 systems in our dataset that disclose a power figure.
Largest powerKw on a system row, among the 523 rows that state one. Power is as reported by the operator or the ranking; the boundary it measures (compute only, or whole facility) varies by source.
113x
more LINPACK per megawatt
The median system that came online in 2020 or later delivers 113 times the LINPACK performance per megawatt of the median system from before 2012 (26.3 against 0.2 PFlop/s per MW), across our dataset.
Rmax divided by power in MW for systems with a measured result and a stated power figure: the median of 153 systems first operational in 2020 or later against 81 first operational before 2012.
68.3
PFlop/s per MW, the best in the dataset
Otus is the most power-efficient system in our dataset on LINPACK: 68.3 PFlop/s for every megawatt.
Rmax divided by power in MW, over 505 systems that have both a measured result and a stated power figure.
91%
of running systems are liquid-cooled
91% of the running systems in our dataset with a recorded cooling design use direct liquid or immersion cooling (58 of 64).
Running systems (operational or upgraded) with a cooling component edge, classified as liquid or immersion against air or other. 902 running systems have no cooling edge recorded and are left out of the denominator, so this describes systems whose cooling is documented, which skews toward newer machines.
44.80 MW
at National Supercomputing Centre in Shenzhen
National Supercomputing Centre in Shenzhen has the highest combined power draw of any site we track: 44.80 MW across its running systems.
Sum of powerKw over running systems at each site, using only systems that state a figure, so a site with undisclosed machines is understated.
38%
of running Rmax is in USA
USA hosts 38% of the combined LINPACK performance of the running systems in our dataset. China is next with 13%.
Sum of Rmax over running systems, grouped by country. Our dataset is curated, not a census, and some countries' largest machines do not submit results, so treat this as the share of what is publicly measured in our sample.
1345
systems in 74 countries
Compute Atlas tracks 1345 supercomputers across 74 countries, each with a named source and a stated confidence level.
Row count of systems.csv and the distinct country values on it.
79%
of accelerated systems use NVIDIA
570 of the 717 accelerated systems in our dataset (79%) use NVIDIA accelerators.
Systems with an accelerator component edge whose supplier rolls up to NVIDIA today, over all systems with any accelerator edge, across all years and statuses.
54% / 38%
of running accelerated Rmax, NVIDIA / AMD
Of the LINPACK performance running on accelerators in our dataset, 54% is on NVIDIA and 38% on AMD.
Rmax summed over running systems with an accelerator edge, split by the accelerator's owner today. A system with both vendors would count in both, so the two can total slightly over 100%.
14%
of supplier links point to a company since acquired
14% of the supplier relationships in our dataset (708 of 4931) are with a company that has since been bought.
Component edges whose supplier has a parent company recorded in companies.csv, over all component edges. The parent is the owner today; the supplier named is the one at build time.
28%
of systems were built by Hewlett Packard Enterprise
Hewlett Packard Enterprise, counting the companies it has acquired, built 332 of the 1189 systems in our dataset with a recorded integrator (28%). The next, Dell Technologies, built 118.
Systems with an integrator component edge, grouped by the integrator's owner today, so Cray and SGI count toward HPE. Systems with no integrator recorded are excluded.
60%
of systems contain a NVIDIA part
NVIDIA supplies at least one component to 809 of the 1345 systems in our dataset (60%), more than any other company.
Distinct systems with at least one component edge whose supplier rolls up to the company today, in any role.
7 months
median from announcement to first operation
The median supercomputer in our dataset took 7 months from announcement to first operation. The fastest, Albert2, took under a month. The slowest, Aurora, took 107.
Months between announcedDate and firstOperationalDate for 221 systems that record both.
6 years
median service life of a retired system
The median retired supercomputer in our dataset ran for 6 years. The longest-serving, Pleiades, ran for 18.
Whole years between first operation and decommissioning, over 144 decommissioned systems that record both. Systems still running are excluded, which biases this figure low.
20%
of systems are verified by two independent sources
Only 20% of the systems in our dataset (271 of 1345) clear our verified bar: two independent organisations naming the same specific fact. The rest are marked reported or estimated.
Count of systems with confidence = verified. Verified requires two independent organisations; a page that copies a press release is not independent.
51
open conflicts between sources
Across 41 systems, we have logged 51 facts where two credible sources give different answers, with both sides linked. We leave the field blank rather than pick one.
Open rows on the open-questions board. Each has two sources that disagree on one specific field.
Found a number that looks wrong? Every fact links to the rows it is built from, and the open-questions board lists where our own sources disagree.