Supercomputers/Analysis/Data trust
Analysis 12
How far to trust this data
234 of 1153 systems (20%) are marked verified, meaning two independent organisations directly confirm the same specific fact. The rest are reported: one credible source, cited, and not contradicted. Nothing in the dataset is guessed, but most of it rests on a single source, and this page says where. When every cited source was re-read, 1092 systems had at least one that names the system, and only 40 had readable sources that all failed to.
What the tiers mean. Verified is a claim about independence: two different organisations, not two pages from the same one. Reported is a claim about provenance: a real source, read, that says this. Estimated is ours, and is used only where the method is published. A reported row is not a doubtful row. It is one whose second source we have not yet found, and the house rule is to publish that rather than to omit it or to upgrade it.
- Verified
- Reported or estimated
Table view: How well each sector is sourced
| Sector | Systems | Verified | Reported | Estimated | Share verified |
|---|---|---|---|---|---|
| Government labs | 233 | 42 | 191 | 0 | 18% |
| Research HPC | 568 | 127 | 441 | 0 | 22% |
| Big tech | 99 | 20 | 79 | 0 | 20% |
| AI labs | 19 | 4 | 15 | 0 | 21% |
| Neoclouds | 85 | 14 | 71 | 0 | 16% |
| Industry | 125 | 22 | 103 | 0 | 18% |
| Finance | 7 | 2 | 5 | 0 | 29% |
| Life sciences | 17 | 3 | 14 | 0 | 18% |
How this was calculated
Each system carries a confidence tier set when it was added: verified only if two independent organisations confirm the same specific fact and both links are recorded, reported for one credible cited source, estimated for our own bounded figures. The same three tiers are recorded on every component edge, so a system can be reported while one of its parts is verified, and the validator checks each edge for its own second source rather than inheriting the system's.
The source re-check, pnpm verify, fetches every cited source for every system
and asks whether the page names the system and mentions the parts and quantities the row
claims. It is a name-mention signal, not a fact check: a page can name a system and still
disagree with the row, and a paywalled or blocking page tells us nothing either way. It
never changes a tier. It proposes a shortlist and a human decides.
By status, by country
| Status | Systems | Verified | Reported | Estimated | Share verified |
|---|---|---|---|---|---|
| operational | 809 | 183 | 626 | 0 | 23% |
| decommissioned | 218 | 40 | 178 | 0 | 18% |
| under construction | 62 | 4 | 58 | 0 | 6% |
| planned | 41 | 4 | 37 | 0 | 10% |
| upgraded | 23 | 3 | 20 | 0 | 13% |
| Country | Systems | Verified | Reported | Estimated | Share verified |
|---|---|---|---|---|---|
| USA | 406 | 62 | 344 | 0 | 15% |
| Japan | 106 | 39 | 67 | 0 | 37% |
| Germany | 87 | 25 | 62 | 0 | 29% |
| China | 58 | 8 | 50 | 0 | 14% |
| United Kingdom | 57 | 7 | 50 | 0 | 12% |
| France | 41 | 7 | 34 | 0 | 17% |
| Canada | 28 | 2 | 26 | 0 | 7% |
| South Korea | 28 | 9 | 19 | 0 | 32% |
| Australia | 24 | 3 | 21 | 0 | 13% |
| India | 24 | 8 | 16 | 0 | 33% |
| Italy | 22 | 9 | 13 | 0 | 41% |
| Saudi Arabia | 22 | 4 | 18 | 0 | 18% |
The component graph
The edge table, the part of the dataset that says which supplier's part is in which machine, has 4258 rows: 724 verified (17%), 3532 reported and 2 estimated. Tiers are set edge by edge, so the graph is better sourced in places than the system rows above it.
What the source re-check found
| Result | Systems | What it means |
|---|---|---|
| Corroborated | 1092 | At least one cited source, re-read, names the system. |
| Mixed | 10 | One source does not name it, another could not be read. |
| Unsupported | 40 | Every readable cited source no longer names the system. |
| Unreadable | 11 | No cited source could be read, which tells us nothing either way. |
| Not yet checked | 0 | Added since the last full re-check. |
Of the 1142 systems with at least one readable source, 96% are corroborated. “Unsupported” is evidence about the citation, not about the fact: a page can be rewritten, moved behind a login or trimmed after we cited it.
Ready to be promoted
27 systems are still marked reported although two different hosts each independently name the system, a part and a quantity. That is the shortlist the re-check proposes for a human to look at. A name-mention cannot establish that two sources agree in the sense verified requires, so none is promoted automatically.
| System | Country | Hosts that name it |
|---|---|---|
| AI-Farabium | Kazakhstan | b2b.telecom.kz, top500.org |
| ARIS | Greece | doc.aris.grnet.gr, top500.org |
| ASPIRE 2A+ | Singapore | nsccsg.github.io, top500.org |
| Athena | Poland | oferta-badawcza.agh.edu.pl, top500.org |
| Barbora | Czechia | docs.it4i.cz, it4i.cz |
| Betzy | Norway | documentation.sigma2.no, top500.org |
| Bridges-2 | USA | api.nsf.gov, psc.edu |
| bwUniCluster 3.0 | Germany | nemo.uni-freiburg.de, wiki.bwhpc.de |
| CM-2 | USA | cisl.ucar.edu, en.wikipedia.org |
| Cray X-MP | USA | cisl.ucar.edu, en.wikipedia.org |
| Cray-2 | USA | cs.lbl.gov, en.wikipedia.org |
| DeltaAI | USA | delta.ncsa.illinois.edu, ncsa.illinois.edu |
| Google TPU v4 supercomputer | USA | arxiv.org, cloud.google.com |
| Govorun | Russia | hlit.jinr.ru, rscgroup.ru |
| Guacolda | Chile | en.wikipedia.org, nlhpc.cl |
| IARA | Brazil | sidi.org.br, top500.org |
| Imperial CX3 | United Kingdom | bebeez.eu, icl-rcs-user-guide.readthedocs.io |
| Iridis 5 | United Kingdom | southampton.ac.uk, top500.org |
| Jairu | Brazil | inforchannel.com.br, telesintese.com.br |
| KEK VPP500/80 | Japan | scwww.kek.jp, top500.org |
| Minerva | USA | labs.icahn.mssm.edu, top500.org |
| Naver B200 4K Cluster | South Korea | koreatechtoday.com, koreatimes.co.kr |
| Project Ceiba | USA | aws.amazon.com, nvidianews.nvidia.com |
| Santos Dumont | Brazil | baguete.com.br, top500.org |
| Stanage | United Kingdom | alces-flight.com, docs.hpc.shef.ac.uk |
| Sunrise | United Kingdom | top500.org, ukaea.org |
| Together AI / Hypertec Cloud GB200 Cluster | USA | prnewswire.com, together.ai |
Needs a second look
40 systems have cited sources that could be read and that do not name them. That does not mean the fact is wrong. Some are naming mismatches, where the page describes the same hardware under another name (the last column shows what the check did find), and some are citations that should be replaced:
| System | Country | What the cited pages did show |
|---|---|---|
| Abe | USA | nothing matching the row |
| Abel (UiO) | Norway | Mellanox InfiniBand FDR; Intel Xeon E5-2670 |
| Aire (Leeds) | United Kingdom | NVIDIA L40S; quantity 84 |
| ALPS (NCHC) | Taiwan | Mellanox InfiniBand QDR |
| Baku (HLRS) | Germany | Windows HPC Server 2008 |
| Blue Horizon | USA | nothing matching the row |
| ByteDance MegaScale Cluster | China | nothing matching the row |
| Cab | USA | Intel Xeon E5-2670; Mellanox InfiniBand QDR |
| CARL and EDDY | Germany | Intel Xeon E5-2650 v4; Mellanox InfiniBand FDR |
| Chebyshev | Russia | Mellanox InfiniBand DDR; CentOS |
| Chengdu Intelligent Computing Center | China | nothing matching the row |
| Chongqing AI Innovation Center | China | nothing matching the row |
| CNIC Era | China | Mellanox InfiniBand FDR |
| CSAR Cray T3E-1200E | United Kingdom | nothing matching the row |
| CZI GPU Cluster | USA | NVIDIA H100 SXM5; VAST Data Platform; quantity 1024 |
| Digital Government AI Cluster | Uzbekistan | nothing matching the row |
| DOST Blue Gene/P | Philippines | nothing matching the row |
| Eka | India | nothing matching the row |
| Gaïa (EDF) | France | Omni-Path Architecture |
| Hawk (Cardiff) | United Kingdom | Intel Xeon Gold 6148; quantity 26; quantity 30 |
| IBM Vela | USA | nothing matching the row |
| Jade (CINES) | France | Intel Xeon E5450 |
| Jade (LLNL) | USA | Intel Xeon E5-2695 v4; Omni-Path Architecture; Tri-Lab Operating System Stack |
| Jean (ARL DSRC) | USA | NVIDIA InfiniBand HDR; CentOS |
| Jean Zay V100 | France | NVIDIA Tesla V100; Omni-Path Architecture |
| JSS FX1 | Japan | Mellanox InfiniBand DDR |
| Julich Cray T3E | Germany | nothing matching the row |
| Kay | Ireland | Intel Xeon Gold 6148; NVIDIA Tesla V100; quantity 64 |
| Lynx (Arizona) | USA | nothing matching the row |
| Maia (NAS) | USA | Intel Xeon E5-2670; Mellanox InfiniBand FDR |
| MCR | USA | Quadrics QsNet |
| Neo | USA | NVIDIA H200 SXM; Tri-Lab Operating System Stack; quantity 32 |
| NIPA (Kakao) | South Korea | Intel Xeon 6 6900P (Granite Rapids); NVIDIA InfiniBand NDR (Quantum-2); NVIDIA B200 SXM |
| Puma (Arizona) | USA | nothing matching the row |
| Qingdao AI Computing Center | China | nothing matching the row |
| Raijin | Japan | NEC SX-Aurora TSUBASA Vector Engine; quantity 4320; NVIDIA InfiniBand HDR |
| Sol (ASU) | USA | Rocky Linux |
| TSUBAME1.0 | Japan | nothing matching the row |
| Vayu (NCI) | Australia | Mellanox InfiniBand QDR |
| Zin | USA | Intel Xeon E5-2670; Mellanox InfiniBand QDR |
The read
The honest summary is that this is a reported dataset with a verified minority. About one system in four (20%) and one component edge in four (17%) has a second independent source behind it. Among the sectors with at least fifteen systems the verified share runs from 16% (Neoclouds) to 22% (Research HPC), and the small sectors move on a single row.
The tier is a column, not a badge of quality, so it should be used as a filter: take the verified rows where a wrong number is expensive, the reported rows for scale and direction, and treat any planned system's capacity as a claim. The fastest way to raise the verified share is not more systems but a second source for the ones already here, and the shortlist above is where that work starts.