Hawk: Rpeak
TOP500 and HLRS disagree; the recorded figure is left unchanged until one is confirmed.
Open questions
Where sources disagree
Most datasets pick a winner and move on. When two credible sources give different figures for the same fact, we record both, leave the row unchanged or blank, and list the conflict here with both links. Each one is a place where a reader with the right knowledge, or an operator with a corrected page, could settle it.
51 open. A conflict leaves this list only when a source settles it, never because we prefer one side.
Hawk: Core count
The recorded figure equals 5,632 nodes times 128 cores; no fetched page states the node count.
MeluXina: Rmax (Accelerator Module)
EuroHPC quotes a higher figure that may be a combined or later result; TOP500 is used.
Horizon: Peak performance
Both pages are TACC; neither figure is recorded.
Horizon: Production date
The later page is treated as current, but no date is recorded.
CEA-HE: Grace Hopper node count
CEA's own October 2025 slides give the larger figure; the 477 is kept.
xAI Colossus: First-phase power
It is unclear whether either is facility or IT load, so power is left blank.
Edison: Core count
TOP500's figure is used.
Capella: Operating system
Possibly a benchmark image versus production; no OS edge is recorded.
Stargate Abilene: Installed GPU count
Up to 50,000 GB200 NVL72 per building (Crusoe)Source
About 201,600 B200 or B300 (analyst estimate)Source
Crusoe's wording is ambiguous between racks and GPUs and the estimate names different parts, so no count is recorded.
Microsoft Fairwater (Atlanta): Scale-out network
No interconnect edge is recorded until one source resolves it.
Polaris: Peak performance
Probably different scopes: the announced system versus the run TOP500 measured.
Hunter: Rank in June 2025
TOP500's figure is used.
SuperMUC-NG: Rpeak
The lower figure is the pre-installation plan; TOP500's measured peak is used.
Miyabi: Storage capacity
No capacity is recorded.
Stampede3: Intel GPU count
40 GPUs in 10 Dell XE9640 servers (press release)Source
80 GPUs in 20 nodes of four Max 1550 (docs)Source
Both pages are TACC; no Intel GPU count is recorded.
Stampede3: Omni-Path speed
The press release may describe the target and the docs the state at the time.
Derecho: A100 GPU count
The 382 looks like a transposition; 328 matches 82 nodes of four.
Expanse: Operating system
TOP500 may be stale; no OS edge is recorded.
Anvil: Operating system
No OS edge is recorded.
KISTI-6: Peak performance
A third source gives 588; no peak is recorded.
KISTI-6: Cooling
No cooling edge is recorded.
JUWELS Cluster: Core count
JSC includes later additions; TOP500 is used.
Pangea III: Peak performance
TOP500's figure is used.
SQUID: Start of service
Only the month is recorded.
Meta Research SuperCluster (RSC): InfiniBand link speed
NVIDIA Quantum 1600 Gb/s InfiniBand two-level Clos fabricSource
NVIDIA Quantum 200Gb/s InfiniBand networkSource
Meta and NVIDIA describe the same fabric at different speeds, so no speed is recorded and no fabric part edge is added.
Taiwania 3: Peak performance
QCT also gives 2.7. No peak is recorded.
Magnus: Core count
The iTnews figure may be a transposition. No core count is recorded.
Pengcheng Cloudbrain-II: Start of operation
Phase one launched 29 November 2019 at 100 PFlop/s (Huawei release)Source
Settled in Shenzhen in September 2020 with tests from October 2020 (Shenzhen news)Source
No first-operational date is recorded.
JURECA-DC: Peak performance
Likely different scopes; no peak is recorded.
JURECA-DC: Core count
No core count is recorded.
JURECA-DC: Operating system
TOP500 may be stale; no OS edge is recorded.
Marconi-100: Peak performance
No peak is recorded.
Tera-1000-2: Core count
No core count is recorded.
Joliot-Curie Rome: Core count and peak
TOP500 may describe a subset of nodes; neither is recorded.
Joliot-Curie Rome: Nodes and interconnect
2,292 nodes, InfiniBand HDR100 (CEA Joliot-Curie page)Source
2,286 nodes, EDR InfiniBand pruned fat tree (TGCC documentation)Source
Both pages are CEA; no interconnect edge is recorded.
Shaheen II: Core count
No core count is recorded.
Yellowstone: Core count
No core count is recorded.
Yellowstone: Processor
No CPU edge is recorded.
Yellowstone: Power draw
The two may measure different things; no figure is recorded.
Santos Dumont: Peak performance
18.85 PFlop/s for the whole machine after expansion (LNCC, via Baguete)Source
20.26 PFlop/s for the 2024 entry (TOP500)Source
The TOP500 entry alone exceeds LNCC total; no peak is recorded.
Comet: End of service
end of NSF service March 2021 (SDSC)Source
75-month operating period to July 2021 (NSF award record)Source
A later Globus post says Comet was retired by August 2021; only the year is recorded.
Comet: NSF funding amount
The award grew through supplements; no figure is recorded.
Nebulae: First operational / active date
2010 (TOP500 installation year, matching its June 2010 TOP500 debut)Source
2011 ("active since 2011")Source
No source fetched for this pass reconciles the two; first_operational_date is left blank.
Lengau: Total storage capacity
4 PB Lustre (TOP500 news coverage)Source
5 PB total Dell storage (CHPC/CSIR unveiling release)Source
CHPC/CSIR and TOP500 give different total storage figures for the same 2016 build; no capacity is recorded in system_components.
PARAM Siddhi-AI: Peak double-precision performance
NVIDIA's own release and TOP500's measured Rpeak disagree on the machine's double-precision peak; TOP500's figure is recorded as Rpeak.
Serafin: Peak performance
147.456 TFlop/s, CCAD's current equipment pageSource
156 TFlop/s, as announced at the August 2021 launchSource
The figure announced at Serafin's 2021 inauguration is about 6% higher than CCAD's own current technical page states for the deployed 60-node configuration; no Rpeak is recorded in the systems table.
Athena: Core count
TOP500's figure is used; AGH's 6,144 appears to count only the AMD EPYC CPU cores, not whatever TOP500 aggregates into its own core count.
Narval: GPU count
632 A100 GPUs (Calcul Québec launch announcement)Source
636 GPUs across 80,720 cores (user portal hardware page)Source
The portal total looks like it includes a few service or login nodes outside TOP500's ranked HPL partition of 76,320 cores; no single figure is recorded as verified.
Miztli: Core / processor count
TOP500's core count is used in the systems table; the press figure may be counting sockets rather than cores, or describing a slightly different point in the machine's history.
Mistral Campus AI (Bruyeres-le-Chatel): Accelerator model and count
18,000 NVIDIA GB200 systems, operational (NVIDIA blog, 17 June 2026)Source
13,800 NVIDIA GB300 GPUs, financed by an $830M debt raise for completion around Q2 2026 (CNBC and Data Center Dynamics, March 2026, as reported by Let's Data Science)Source
Both describe the same 44MW Bruyeres-le-Chatel facility but disagree on GPU generation and count; only NVIDIA's more recent figure is recorded on the system row.