Supercomputers/Analysis/The shadow list
Analysis 04
The shadow list
100 clusters that appear on no public ranking anywhere hold 623,166 anchorable accelerators between them. Bracketed against measured systems in this dataset using the same silicon, they come to 17.73 EFlop/s to 72.41 EFlop/s of HPL-equivalent compute, against 18.49 EFlop/s for every ranked system here combined. That is 96 to 392% of the visible world, from 100 machines belonging to 88 operators, each disclosed in the operator's own blog post, release or investor deck.
This is the page we refused to publish three times, and here is what changed. An estimate is only worth publishing if every input is traceable. The estimator below uses no vendor peak figures and no outside numbers: each bound comes from dividing two sourced fields on one of our own system rows. Where that is not possible, this page publishes nothing for that system and says so by name. 59 could not be anchored and carry no estimate.
Phase 1 seed dataset, compiled by hand. These rows were built from public operator, laboratory and vendor sources. A mechanical second-reader pass has since fetched every cited source: 1263 of 1345 systems have a readable citation that names them, and 56 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Read the shape, not the spikes. This is the installed base of the 1345 systems in this dataset: a curated seed chosen for supply-chain diversity, not a census of the market. At that sample size one machine entering or leaving service moves a share line by tens of points, so year-to-year jumps are composition effects rather than market events. The multi-year trends are the part worth quoting; the individual steps usually have one system's name on them, and the table view under the chart will tell you which.
- Ranked systems
- 848
- Ranked total
- 18.49 EFlop/s
- Unranked systems
- 495
- Shadow estimate
- 17.73 EFlop/s to 72.41 EFlop/s
- As a share of ranked
- 96 to 392%
measured performance
351 disclose no hardware at all
estimator anchor-bracket-v1
The measured basis
Every estimate on this page rests on this table and nothing else. These are the systems in the dataset that publish both an Rmax and a single accelerator count, so dividing one by the other gives a measured HPL rate per accelerator. No vendor peak figure appears anywhere in the calculation.
- Intel
- NVIDIA
- AMD
Table view: Measured HPL rate per accelerator
| Accelerator | Measured in | Year | Count | System Rmax | TFlop/s each |
|---|---|---|---|---|---|
| Intel Data Center GPU Max 1550 | Clementina XXI | 0 | 296 | 5.39 PFlop/s | 18.21 |
| Intel Data Center GPU Max 1550 | SuperMUC-NG Phase 2 | 0 | 960 | 17.19 PFlop/s | 17.91 |
| NVIDIA GH200 Grace Hopper Superchip | KAIROS | 0 | 64 | 3.05 PFlop/s | 47.66 |
| NVIDIA H100 80GB PCIe | Pegasus (Tsukuba) | 0 | 150 | 4.34 PFlop/s | 28.93 |
| NVIDIA H100 SXM5 | Snellius (Phase 3 GPU) | 0 | 352 | 13.64 PFlop/s | 38.75 |
| NVIDIA H100 SXM5 | SAKURAONE | 0 | 800 | 33.95 PFlop/s | 42.44 |
| NVIDIA H100 SXM5 | Taipei-1 | 0 | 512 | 22.30 PFlop/s | 43.55 |
| NVIDIA H200 SXM | Redtail | 0 | 264 | 12.85 PFlop/s | 48.67 |
| NVIDIA Tesla V100 | Taiwania 2 | 0 | 2,016 | 9.00 PFlop/s | 4.46 |
| NVIDIA Tesla V100 | Alvis | 0 | 44 | 1.94 PFlop/s | 44.09 |
| ATI Radeon HD 4870 X2 | Tianhe-1 | 2009 | 2,560 | 563.10 TFlop/s | 0.22 |
| NVIDIA Tesla M2050 | Tianhe-1A | 2010 | 7,168 | 2.57 PFlop/s | 0.36 |
| NVIDIA Tesla M2050 | TSUBAME2.0 | 2010 | 4,224 | 1.19 PFlop/s | 0.28 |
| NVIDIA Tesla K20X | Todi | 2011 | 272 | 273.70 TFlop/s | 1.01 |
| NVIDIA Tesla K20X | Titan | 2012 | 18,688 | 17.59 PFlop/s | 0.94 |
| Intel Xeon Phi SE10P | Stampede | 2013 | 6,880 | 5.17 PFlop/s | 0.75 |
| NVIDIA Tesla K20X | TSUBAME-KFC | 2013 | 160 | 125.20 TFlop/s | 0.78 |
| NVIDIA Tesla K20X | Lichtenberg I | 2013 | 86 | 213.78 TFlop/s | 2.49 |
| NVIDIA Tesla K20X | Hydra (MPCDF) | 2013 | 676 | 1.28 PFlop/s | 1.90 |
| AMD FirePro S9150 | L-CSC | 2014 | 640 | 316.70 TFlop/s | 0.49 |
| Intel Xeon Phi 31S1P | Tianhe-2 LvLiang Solution | 2014 | 2,688 | 2.07 PFlop/s | 0.77 |
| NVIDIA Tesla V100 | ARIS | 2015 | 8 | 170.00 TFlop/s | 21.25 |
| Intel Xeon Phi 7250 (Knights Landing) | Camphor 2 | 2016 | 1,800 | 3.06 PFlop/s | 1.70 |
| NVIDIA Tesla P100 | DGX SaturnV | 2016 | 992 | 4.00 PFlop/s | 4.03 |
| NVIDIA Tesla P100 | Owens | 2016 | 160 | 705.97 TFlop/s | 4.41 |
| NVIDIA Tesla P100 | Reedbush-H | 2017 | 240 | 800.00 TFlop/s | 3.33 |
| NVIDIA Tesla P100 | Reedbush-L | 2017 | 256 | 810.00 TFlop/s | 3.16 |
| NVIDIA Tesla P100 | Cedar (SFU) | 2017 | 584 | 1.34 PFlop/s | 2.29 |
| NVIDIA Tesla P100 | AIST AI Cloud | 2017 | 400 | 961.00 TFlop/s | 2.40 |
| NVIDIA Tesla V100 | Iridis 5 | 2017 | 20 | 860.00 TFlop/s | 43.00 |
| NVIDIA Tesla V100 | Raad2 | 2017 | 8 | 120.00 TFlop/s | 15.00 |
| PEZY-SC2 | Gyoukou | 2017 | 10,000 | 19.14 PFlop/s | 1.91 |
| NVIDIA Tesla P100 | HPC4 | 2018 | 3,200 | 12.21 PFlop/s | 3.82 |
| NVIDIA Tesla P100 | Taiwania | 2018 | 256 | 1.33 PFlop/s | 5.18 |
| NVIDIA Tesla V100 | Sierra | 2018 | 17,280 | 94.64 PFlop/s | 5.48 |
| NVIDIA Tesla V100 | Summit | 2018 | 27,648 | 148.6 PFlop/s | 5.37 |
| NVIDIA Tesla V100 | Lassen | 2018 | 3,168 | 18.20 PFlop/s | 5.74 |
| NVIDIA Tesla V100 | JUWELS Cluster | 2018 | 224 | 6.18 PFlop/s | 27.59 |
| NVIDIA Tesla V100 | Kay | 2018 | 64 | 665.00 TFlop/s | 10.39 |
| NVIDIA Tesla V100 | COBRA | 2018 | 128 | 5.61 PFlop/s | 43.83 |
| NVIDIA Tesla V100 | CLAIX-2018 | 2018 | 96 | 2.48 PFlop/s | 25.83 |
| NVIDIA Tesla V100 | MareNostrum P9 CTE | 2018 | 216 | 1.02 PFlop/s | 4.71 |
| NVIDIA Tesla V100 | Circe | 2018 | 576 | 3.06 PFlop/s | 5.31 |
| NVIDIA Tesla V100 | TX-GAIA | 2019 | 896 | 5.16 PFlop/s | 5.76 |
| NVIDIA Tesla V100 | Cygnus | 2019 | 320 | 1.58 PFlop/s | 4.94 |
| NVIDIA Tesla V100 | Jean Zay V100 | 2019 | 1,044 | 4.48 PFlop/s | 4.29 |
| NVIDIA Tesla V100 | Cannon (FASRC) | 2019 | 64 | 2.08 PFlop/s | 32.50 |
| NVIDIA Tesla V100 | Béluga | 2019 | 688 | 2.28 PFlop/s | 3.31 |
| NVIDIA Tesla V100 | Longhorn | 2019 | 448 | 2.29 PFlop/s | 5.11 |
| NVIDIA Tesla V100 | NVIDIA DGX SuperPOD (V100) | 2019 | 1,536 | 9.44 PFlop/s | 6.15 |
| NEC SX-Aurora TSUBASA Vector Engine | Raijin | 2020 | 4,320 | 7.89 PFlop/s | 1.83 |
| NVIDIA A100 SXM4 40GB | Betzy | 2020 | 16 | 4.72 PFlop/s | 295.00 |
| NVIDIA A100 SXM4 40GB | JURECA-DC | 2020 | 768 | 9.33 PFlop/s | 12.15 |
| NVIDIA A100 SXM4 40GB | PARAM Siddhi-AI | 2020 | 336 | 4.62 PFlop/s | 13.75 |
| NVIDIA A100 SXM4 40GB | Lyapunov | 2020 | 1,096 | 12.81 PFlop/s | 11.69 |
| NVIDIA Tesla V100 | HPC5 | 2020 | 7,280 | 35.45 PFlop/s | 4.87 |
| NVIDIA Tesla V100 | Expanse | 2020 | 208 | 2.48 PFlop/s | 11.92 |
| NVIDIA Tesla V100 | Marconi-100 | 2020 | 3,920 | 21.64 PFlop/s | 5.52 |
| NVIDIA Tesla V100 | JADE2 | 2020 | 504 | 1.47 PFlop/s | 2.92 |
| Preferred Networks MN-Core (1st generation) | MN-3 | 2020 | 128 | 2.18 PFlop/s | 17.03 |
| NVIDIA A100 (in DGX A100) | Cambridge-1 | 2021 | 640 | 9.68 PFlop/s | 15.13 |
| NVIDIA A100 (in DGX A100) | IARA | 2021 | 25 | 3.66 PFlop/s | 146.40 |
| NVIDIA A100 SXM4 40GB | SQUID | 2021 | 336 | 6.11 PFlop/s | 18.18 |
| NVIDIA A100 SXM4 40GB | Athena | 2021 | 384 | 5.05 PFlop/s | 13.15 |
| NVIDIA A100 SXM4 40GB | Narval | 2021 | 632 | 5.89 PFlop/s | 9.32 |
| NVIDIA A100 SXM4 40GB | Raven-GPU | 2021 | 768 | 8.62 PFlop/s | 11.22 |
| NVIDIA A100 SXM4 40GB | Baskerville | 2021 | 184 | 2.34 PFlop/s | 12.72 |
| NVIDIA B200 SXM | HiPerGator AI | 2021 | 504 | 17.49 PFlop/s | 34.70 |
| NVIDIA H200 SXM | Discoverer | 2021 | 32 | 4.52 PFlop/s | 141.25 |
| NVIDIA H200 SXM | HoreKa | 2021 | 60 | 8.03 PFlop/s | 133.83 |
| NVIDIA Tesla V100 | Altair | 2021 | 72 | 3.53 PFlop/s | 49.03 |
| NVIDIA Tesla V100 | Ares | 2021 | 72 | 2.34 PFlop/s | 32.50 |
| AMD Instinct MI250X | Frontier | 2022 | 37,632 | 1.35 EFlop/s | 35.95 |
| AMD Instinct MI250X | LUMI | 2022 | 11,912 | 379.7 PFlop/s | 31.88 |
| NVIDIA A100 SXM4 40GB | Lucia | 2022 | 200 | 2.72 PFlop/s | 13.60 |
| NVIDIA A100 SXM4 40GB | VSC-5 | 2022 | 120 | 2.31 PFlop/s | 19.25 |
| NVIDIA A100 SXM4 40GB | LANTA | 2022 | 704 | 8.15 PFlop/s | 11.58 |
| NVIDIA A100 SXM4 64GB | Leonardo | 2022 | 13,824 | 241.2 PFlop/s | 17.45 |
| NVIDIA A100 SXM4 80GB | SK Telecom Titan | 2022 | 1,040 | 19.53 PFlop/s | 18.78 |
| NVIDIA A100 SXM4 80GB | Champollion | 2022 | 160 | 2.32 PFlop/s | 14.50 |
| NVIDIA H100 80GB PCIe | Henri | 2022 | 80 | 2.88 PFlop/s | 36.00 |
| Graphcore IPU | fortytwo | 2023 | 4 | 7.94 PFlop/s | 1985.00 |
| Intel Data Center GPU Max 1550 | Aurora | 2023 | 63,744 | 1.01 EFlop/s | 15.88 |
| NEC SX-Aurora TSUBASA Vector Engine | AOBA-S | 2023 | 4,032 | 17.22 PFlop/s | 4.27 |
| NVIDIA A100 SXM4 40GB | Derecho | 2023 | 328 | 10.32 PFlop/s | 31.46 |
| NVIDIA A100 SXM4 40GB | ASPIRE 2A (GPU Partition) | 2023 | 352 | 3.33 PFlop/s | 9.46 |
| NVIDIA H100 SXM5 | Israel-1 | 2023 | 2,048 | 41.50 PFlop/s | 20.26 |
| NVIDIA H100 SXM5 | Pre-Eos | 2023 | 1,024 | 40.66 PFlop/s | 39.71 |
| AMD Instinct MI300A | Blueback | 2024 | 128 | 10.30 PFlop/s | 80.47 |
| Intel Data Center GPU Max 1550 | Dawn | 2024 | 1,024 | 19.46 PFlop/s | 19.00 |
| NVIDIA B200 SXM | Minerva | 2024 | 48 | 7.91 PFlop/s | 164.79 |
| NVIDIA GH200 Grace Hopper Superchip | Alps | 2024 | 10,752 | 434.9 PFlop/s | 40.45 |
| NVIDIA GH200 Grace Hopper Superchip | Helios | 2024 | 440 | 19.14 PFlop/s | 43.50 |
| NVIDIA GH200 Grace Hopper Superchip | ROMEO-2025 | 2024 | 232 | 9.86 PFlop/s | 42.50 |
| NVIDIA H100 80GB PCIe | amplitUDE | 2024 | 72 | 1.95 PFlop/s | 27.08 |
| NVIDIA H100 SXM5 | ASPIRE 2A+ | 2024 | 320 | 14.20 PFlop/s | 44.38 |
| NVIDIA H100 SXM5 | BioHive-2 | 2024 | 504 | 23.32 PFlop/s | 46.27 |
| NVIDIA H100 SXM5 | Gautschi AI | 2024 | 160 | 7.26 PFlop/s | 45.38 |
| NVIDIA H100 SXM5 | Ubilink H100 Cluster | 2024 | 1,024 | 45.82 PFlop/s | 44.75 |
| NVIDIA H100 SXM5 | fortythree | 2024 | 280 | 11.27 PFlop/s | 40.25 |
| NVIDIA H100 SXM5 94GB | MUSICA | 2024 | 1,088 | 31.84 PFlop/s | 29.26 |
| NVIDIA H100 SXM5 94GB | PITAGORA | 2024 | 672 | 27.90 PFlop/s | 41.52 |
| NVIDIA H100 SXM5 94GB | Kuma | 2024 | 336 | 12.31 PFlop/s | 36.64 |
| NVIDIA H100 SXM5 94GB | Cardinal | 2024 | 128 | 5.04 PFlop/s | 39.38 |
| NVIDIA H100 SXM5 94GB | Lem | 2024 | 304 | 12.80 PFlop/s | 42.11 |
| NVIDIA H100 SXM5 94GB | Kraken-Fregata | 2024 | 148 | 5.98 PFlop/s | 40.41 |
| NVIDIA H200 SXM | ARF-ACC | 2024 | 192 | 12.09 PFlop/s | 62.97 |
| NVIDIA H200 SXM | GMO GPU Cloud | 2024 | 256 | 10.05 PFlop/s | 39.26 |
| AMD Instinct MI300A | Hunter | 2025 | 752 | 31.68 PFlop/s | 42.13 |
| AMD Instinct MI300A | Viper-GPU | 2025 | 654 | 31.10 PFlop/s | 47.55 |
| AMD Instinct MI300A | Cosmos (SDSC) | 2025 | 168 | 5.18 PFlop/s | 30.83 |
| NVIDIA B200 SXM | SEIRAN | 2025 | 96 | 3.32 PFlop/s | 34.58 |
| NVIDIA B200 SXM | Betty | 2025 | 248 | 8.49 PFlop/s | 34.23 |
| NVIDIA GH200 Grace Hopper Superchip | Miyabi | 2025 | 1,120 | 46.80 PFlop/s | 41.79 |
| NVIDIA GH200 Grace Hopper Superchip | Olivia | 2025 | 448 | 13.20 PFlop/s | 29.46 |
| NVIDIA H100 SXM5 | Space HPC | 2025 | 108 | 4.47 PFlop/s | 41.39 |
| NVIDIA H100 SXM5 | Trillium | 2025 | 252 | 9.49 PFlop/s | 37.66 |
| NVIDIA H100 SXM5 | Otus | 2025 | 108 | 4.66 PFlop/s | 43.15 |
| NVIDIA H100 SXM5 | Rorqual | 2025 | 324 | 6.97 PFlop/s | 21.51 |
| NVIDIA H100 SXM5 | Matrix | 2025 | 112 | 3.11 PFlop/s | 27.77 |
| NVIDIA H200 SXM | Torch | 2025 | 272 | 10.79 PFlop/s | 39.67 |
| NVIDIA H200 SXM | Cannon2-H200 | 2025 | 96 | 4.24 PFlop/s | 44.17 |
| NVIDIA H200 SXM | PERUN | 2025 | 208 | 10.21 PFlop/s | 49.09 |
| NVIDIA H200 SXM | AI-Farabium | 2025 | 400 | 17.93 PFlop/s | 44.83 |
| AMD Instinct MI300A | Sirius (Tsukuba) | 2026 | 96 | 5.63 PFlop/s | 58.65 |
| AMD Instinct MI355X | Sunrise | 2026 | 672 | 34.77 PFlop/s | 51.74 |
How this was calculated, in full
Step 1. For every system publishing both an Rmax and exactly one accelerator count, divide the first by the second. Systems with more than one accelerator supplier are excluded, because the split of their Rmax between parts is not observable.
Step 2. For an unranked system with a disclosed accelerator count, bracket it with the lowest and highest measured rate for that same part. Where a part appears in several measured systems, the spread between them is the range. Frontier and LUMI both run MI250X and differ by about 12%; Summit and Sierra both run V100 and differ by about 2%. That spread is real and it is what a range should represent.
Step 3. Where a part has only one measured observation, a point estimate would be false precision, so it is widened by 15% either way, taken from the largest observed spread in step 2.
Step 4. Where no measured rate exists for the part and no stated proxy applies, publish nothing for that system.
The one substitution, stated rather than hidden. This dataset contains no measured HPL rate for the H100, because no system here publishes both an H100 count and an Rmax. It does contain a measured rate for the GH200 superchip, from Alps. GH200 is a Hopper GPU packaged with a Grace CPU. HPL runs on the GPU, so the measured GH200 rate is close to a measured H100 rate. That substitution is the single weakest link in this page, and if you disagree with it, the number to change is in the CSV.
What this estimator refuses to do. It does not multiply a vendor peak figure by an assumed efficiency. That method needs a peak number, an efficiency assumption, a utilisation assumption and a hardware mix assumption, and four uncertain inputs multiplied together produce a number with no meaningful confidence interval that nonetheless gets quoted as fact.
The estimate
| System | Operator | Accelerator | Count | HPL-equivalent | Anchored on |
|---|---|---|---|---|---|
| Meta GenAI cluster (RoCE) | Meta Platforms | NVIDIA H100 SXM5 | 24,576 | 498.0 PFlop/s to 1.14 EFlop/s | Israel-1 |
| Meta GenAI cluster (InfiniBand) | Meta Platforms | NVIDIA H100 SXM5 | 24,576 | 498.0 PFlop/s to 1.14 EFlop/s | Israel-1 |
| xAI Colossus | xAI | NVIDIA H100 SXM5 | 100,000 | 2.03 EFlop/s to 4.63 EFlop/s | Israel-1 |
| Tesla Cortex | Tesla | NVIDIA H100 SXM5 | 50,000 | 1.01 EFlop/s to 2.31 EFlop/s | Israel-1 |
| Tesla Cortex | Tesla | NVIDIA H200 SXM | 16,000 | 628.1 PFlop/s to 2.26 EFlop/s | GMO GPU Cloud |
| Google TPU v5p pod | Alphabet | Google TPU v5p | 8,960 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google TPU7x (Ironwood) pod | Alphabet | Google TPU7x (Ironwood) | 9,216 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Solstice | Argonne National Laboratory | NVIDIA B200 SXM | 100,000 | 3.42 EFlop/s to 16.48 EFlop/s | Betty |
| Equinox | Argonne National Laboratory | NVIDIA B200 SXM | 10,000 | 342.3 PFlop/s to 1.65 EFlop/s | Betty |
| Condor Galaxy 1 | Cerebras | Cerebras CS-2 (WSE-2) | 64 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Stampede3 | Texas Advanced Computing Center | NVIDIA H100 SXM5 | 96 | 1.95 PFlop/s to 4.44 PFlop/s | Israel-1 |
| Bridges-2 | PSC | NVIDIA Tesla V100 | 280 | 816.67 TFlop/s to 13.73 PFlop/s | JADE2 |
| Bridges-2 | PSC | NVIDIA H100 SXM5 | 80 | 1.62 PFlop/s to 3.70 PFlop/s | Israel-1 |
| Meta Research SuperCluster (RSC) | Meta Platforms | NVIDIA A100 (in DGX A100) | 6,080 | 91.96 PFlop/s to 890.1 PFlop/s | Cambridge-1 |
| Google TPU v4 supercomputer | Alphabet | Google TPU v4 | 4,096 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Cerebras Andromeda | Cerebras | Cerebras CS-2 (WSE-2) | 16 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Condor Galaxy 3 | Cerebras | Cerebras WSE-3 | 64 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Comet | SDSC | NVIDIA Tesla P100 | 144 | 330.41 TFlop/s to 745.40 TFlop/s | Cedar (SFU) |
| CoreWeave AI cluster for Inflection AI | CoreWeave | NVIDIA H100 SXM5 | 3,584 | 72.63 PFlop/s to 165.8 PFlop/s | Israel-1 |
| GroqCloud Dammam Inference Cluster | Groq | Groq LPU | 19,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Microsoft Azure AI supercomputer (Loughton, UK) | Microsoft | NVIDIA GB300 NVL72 | 23,040 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Microsoft Azure AI datacenter (Narvik, Norway) | Microsoft | NVIDIA Vera Rubin | 30,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Topaz | Pawsey Supercomputing Research Centre | NVIDIA Tesla V100 | 44 | 128.33 TFlop/s to 2.16 PFlop/s | JADE2 |
| Bridges | PSC | NVIDIA Tesla V100 | 88 | 256.67 TFlop/s to 4.31 PFlop/s | JADE2 |
| Mistral Campus AI (Bruyeres-le-Chatel) | Mistral AI | NVIDIA GB200 NVL72 | 18,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Together AI / Hypertec Cloud GB200 Cluster | Together AI | NVIDIA GB200 NVL72 | 36,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Databricks DBRX Training Cluster | Databricks | NVIDIA H100 SXM5 | 3,072 | 62.25 PFlop/s to 142.1 PFlop/s | Israel-1 |
| Sunrise | UKAEA | AMD Instinct MI355X | 672 | 29.55 PFlop/s to 39.99 PFlop/s | Sunrise |
| Kay | ICHEC | NVIDIA Tesla V100 | 64 | 186.67 TFlop/s to 3.14 PFlop/s | JADE2 |
| Euler | ETH Zurich | NVIDIA A100 SXM4 40GB | 3 | 27.96 TFlop/s to 885.00 TFlop/s | Narval |
| Della | Princeton University | NVIDIA H100 SXM5 | 336 | 6.81 PFlop/s to 15.55 PFlop/s | Israel-1 |
| Anton 2 | D. E. Shaw Research | Anton 2 ASIC | 512 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Anton 3 | D. E. Shaw Research | Anton 3 ASIC | 512 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Cyclone | The Cyprus Institute | NVIDIA Tesla V100 | 64 | 186.67 TFlop/s to 3.14 PFlop/s | JADE2 |
| Rudens | Riga Technical University | NVIDIA Tesla V100 | 8 | 23.33 TFlop/s to 392.22 TFlop/s | JADE2 |
| farm22 | Wellcome Sanger Institute | NVIDIA H100 SXM5 | 64 | 1.30 PFlop/s to 2.96 PFlop/s | Israel-1 |
| farm22 | Wellcome Sanger Institute | NVIDIA H200 SXM | 80 | 3.14 PFlop/s to 11.30 PFlop/s | GMO GPU Cloud |
| farm22 | Wellcome Sanger Institute | NVIDIA L40S | 40 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| farm22 | Wellcome Sanger Institute | NVIDIA Tesla V100 | 30 | 87.50 TFlop/s to 1.47 PFlop/s | JADE2 |
| farm22 | Wellcome Sanger Institute | NVIDIA GH200 Grace Hopper Superchip | 4 | 117.86 TFlop/s to 190.63 TFlop/s | Olivia |
| SINES Data Campus GB300 Deployment | Microsoft | NVIDIA GB300 NVL72 | 12,600 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| PARAM Pravega | IISc | NVIDIA Tesla V100 | 80 | 233.33 TFlop/s to 3.92 PFlop/s | JADE2 |
| Tesla Dojo | Tesla | Tesla Dojo D1 | 3,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Firebird Armenian AI Factory (DC-1) | Firebird | NVIDIA B200 SXM | 6,144 | 210.3 PFlop/s to 1.01 EFlop/s | Betty |
| Haein Cluster | SK Telecom | NVIDIA B200 SXM | 1,000 | 34.23 PFlop/s to 164.8 PFlop/s | Betty |
| Industrial AI Cloud | Deutsche Telekom | NVIDIA B200 SXM | 10,000 | 342.3 PFlop/s to 1.65 EFlop/s | Betty |
| NeXXt AI Factory | Fastweb | NVIDIA H100 SXM5 | 248 | 5.03 PFlop/s to 11.47 PFlop/s | Israel-1 |
| Mayo Clinic DGX SuperPOD | Mayo Clinic | NVIDIA B200 SXM | 128 | 4.38 PFlop/s to 21.09 PFlop/s | Betty |
| St. Jude HPC Research Cluster | St. Jude Children's Research Hospital | NVIDIA H100 SXM5 | 112 | 2.27 PFlop/s to 5.18 PFlop/s | Israel-1 |
| Naver B200 4K Cluster | Naver | NVIDIA B200 SXM | 4,000 | 136.9 PFlop/s to 659.2 PFlop/s | Betty |
| TensorWave Tucson MI325X Cluster | TensorWave | AMD Instinct MI325X | 8,192 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| wICE | KU Leuven | NVIDIA H100 SXM5 | 20 | 405.27 TFlop/s to 925.40 TFlop/s | Israel-1 |
| Hon Hai Kaohsiung Super Computing Center | Hon Hai Technology Group (Foxconn) | NVIDIA GB200 NVL72 | 64 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Leftraru Epu | University of Chile | AMD Instinct MI210 | 12 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Rocket Cluster | University of Tartu | NVIDIA Tesla V100 | 44 | 128.33 TFlop/s to 2.16 PFlop/s | JADE2 |
| Serbia National AI Platform | Office for IT and eGovernment of Serbia | NVIDIA A100 (in DGX A100) | 32 | 484.00 TFlop/s to 4.68 PFlop/s | Cambridge-1 |
| Serbia National AI Platform | Office for IT and eGovernment of Serbia | NVIDIA H200 SXM | 48 | 1.88 PFlop/s to 6.78 PFlop/s | GMO GPU Cloud |
| HPC Saulėtekis | Vilnius University | NVIDIA Tesla V100 | 32 | 93.33 TFlop/s to 1.57 PFlop/s | JADE2 |
| Tokyo-1 | Xeureka | NVIDIA H100 SXM5 | 128 | 2.59 PFlop/s to 5.92 PFlop/s | Israel-1 |
| Freyja | Amgen | NVIDIA H100 SXM5 | 248 | 5.03 PFlop/s to 11.47 PFlop/s | Israel-1 |
| CZI GPU Cluster | Chan Zuckerberg Initiative | NVIDIA H100 SXM5 | 1,024 | 20.75 PFlop/s to 47.38 PFlop/s | Israel-1 |
| Kyber | InstaDeep | NVIDIA H100 SXM5 | 224 | 4.54 PFlop/s to 10.36 PFlop/s | Israel-1 |
| Roihu | CSC | NVIDIA GH200 Grace Hopper Superchip | 528 | 15.56 PFlop/s to 25.16 PFlop/s | Olivia |
| Ares | Cyfronet AGH | NVIDIA Tesla V100 | 72 | 210.00 TFlop/s to 3.53 PFlop/s | JADE2 |
| Plasma Simulator Sosei | NIFS | AMD Instinct MI300A | 280 | 8.63 PFlop/s to 22.53 PFlop/s | Cosmos (SDSC) |
| Project Ceiba | Amazon | NVIDIA B200 SXM | 20,736 | 709.9 PFlop/s to 3.42 EFlop/s | Betty |
| IBM Blue Vela | IBM | NVIDIA H100 SXM5 | 5,000 | 101.3 PFlop/s to 231.3 PFlop/s | Israel-1 |
| Google TPU v4 ML Hub (Oklahoma) | Alphabet | Google TPU v4 | 32,768 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Tinkercliffs | Virginia Tech | NVIDIA H200 SXM | 48 | 1.88 PFlop/s to 6.78 PFlop/s | GMO GPU Cloud |
| Nabuchodonosor | Scaleway | NVIDIA H100 SXM5 | 1,016 | 20.59 PFlop/s to 47.01 PFlop/s | Israel-1 |
| Taiwan AI Factory | GMI Cloud | NVIDIA GB300 NVL72 | 7,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Killarney | Vector Institute | NVIDIA L40S | 672 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Killarney | Vector Institute | NVIDIA H100 SXM5 | 80 | 1.62 PFlop/s to 3.70 PFlop/s | Israel-1 |
| Vulcan (Amii) | Amii | NVIDIA L40S | 820 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Bede | N8 CIR | NVIDIA Tesla V100 | 128 | 373.33 TFlop/s to 6.28 PFlop/s | JADE2 |
| Bede | N8 CIR | NVIDIA GH200 Grace Hopper Superchip | 9 | 265.18 TFlop/s to 428.91 TFlop/s | Olivia |
| MN-2 | Preferred Networks | NVIDIA Tesla V100 | 1,024 | 2.99 PFlop/s to 50.20 PFlop/s | JADE2 |
| Neocortex | PSC | Cerebras WSE-3 | 2 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Vista (TACC) | Texas Advanced Computing Center | NVIDIA GH200 Grace Hopper Superchip | 600 | 17.68 PFlop/s to 28.59 PFlop/s | Olivia |
| Jetstream2 | Indiana University | NVIDIA A100 SXM4 40GB | 360 | 3.36 PFlop/s to 106.2 PFlop/s | Narval |
| Jetstream2 | Indiana University | NVIDIA H100 SXM5 | 96 | 1.95 PFlop/s to 4.44 PFlop/s | Israel-1 |
| Jetstream2 | Indiana University | NVIDIA L40S | 32 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| SDS-AI | SDS Shonfeld | NVIDIA B200 SXM | 2,032 | 69.56 PFlop/s to 334.9 PFlop/s | Betty |
| Nebius Modiin Cluster | Nebius | NVIDIA B200 SXM | 4,000 | 136.9 PFlop/s to 659.2 PFlop/s | Betty |
| PARAM Shakti | IIT Kharagpur | NVIDIA Tesla V100 | 44 | 128.33 TFlop/s to 2.16 PFlop/s | JADE2 |
| E2E Chennai B200 Cluster | E2E Networks | NVIDIA B200 SXM | 1,024 | 35.06 PFlop/s to 168.7 PFlop/s | Betty |
| Govorun | JINR | NVIDIA H100 80GB PCIe | 16 | 433.33 TFlop/s to 576.00 TFlop/s | amplitUDE |
| Govorun | JINR | NVIDIA Tesla V100 | 40 | 116.67 TFlop/s to 1.96 PFlop/s | JADE2 |
| Zhores | Skoltech | NVIDIA Tesla V100 | 104 | 303.33 TFlop/s to 5.10 PFlop/s | JADE2 |
| NASA Cabeus | NASA | NVIDIA GH200 Grace Hopper Superchip | 350 | 10.31 PFlop/s to 16.68 PFlop/s | Olivia |
| SCINet Atlas | USDA ARS | NVIDIA Tesla V100 | 8 | 23.33 TFlop/s to 392.22 TFlop/s | JADE2 |
| SCINet Atlas | USDA ARS | NVIDIA L40S | 48 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Cerebras Oklahoma City | Cerebras | Cerebras WSE-3 | 300 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| LRZ Cerebras CS-2 | LRZ | Cerebras CS-2 (WSE-2) | 1 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| EPCC Cerebras CS-2 Cluster | EPCC | Cerebras CS-2 (WSE-2) | 2 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| HOLL-I | NCSA | Cerebras CS-2 (WSE-2) | 1 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Herder | HLRS | AMD Instinct MI430X | 1,488 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| rzVernal | Lawrence Livermore National Laboratory | AMD Instinct MI250X | 152 | 4.85 PFlop/s to 5.46 PFlop/s | LUMI |
| Crusher | Oak Ridge National Laboratory | AMD Instinct MI250X | 768 | 24.48 PFlop/s to 27.61 PFlop/s | LUMI |
| mdx I | University of Tokyo | NVIDIA A100 SXM4 40GB | 320 | 2.98 PFlop/s to 94.40 PFlop/s | Narval |
| Marlowe (Stanford) | Stanford University | NVIDIA H100 SXM5 | 248 | 5.03 PFlop/s to 11.47 PFlop/s | Israel-1 |
| Seren (Shanghai AI Lab) | Shanghai AI Laboratory | NVIDIA A100 SXM4 80GB | 2,288 | 33.18 PFlop/s to 42.97 PFlop/s | Champollion |
| Kalos (Shanghai AI Lab) | Shanghai AI Laboratory | NVIDIA A100 SXM4 80GB | 2,416 | 35.03 PFlop/s to 45.37 PFlop/s | Champollion |
| Triton | Aalto University | NVIDIA H200 SXM | 120 | 4.71 PFlop/s to 16.95 PFlop/s | GMO GPU Cloud |
| Triton | Aalto University | NVIDIA Tesla V100 | 288 | 840.00 TFlop/s to 14.12 PFlop/s | JADE2 |
| Triton | Aalto University | AMD Instinct MI210 | 2 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Pelle | UPPMAX | NVIDIA L40S | 40 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Imperial HX1 | Imperial College London | NVIDIA A100 SXM4 80GB | 60 | 870.00 TFlop/s to 1.13 PFlop/s | Champollion |
| Imperial CX3 | Imperial College London | NVIDIA L40S | 56 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Kelvin2 | NI-HPC | NVIDIA Tesla V100 | 32 | 93.33 TFlop/s to 1.57 PFlop/s | JADE2 |
| Computerome 2.0 | Technical University of Denmark | NVIDIA Tesla V100 | 40 | 116.67 TFlop/s to 1.96 PFlop/s | JADE2 |
| DARWIN | University of Delaware | NVIDIA Tesla V100 | 12 | 35.00 TFlop/s to 588.33 TFlop/s | JADE2 |
| Redtail | University of Utah | NVIDIA H200 SXM | 264 | 10.36 PFlop/s to 37.29 PFlop/s | GMO GPU Cloud |
| Resnick HPC Cluster | California Institute of Technology | NVIDIA Tesla P100 | 200 | 458.90 TFlop/s to 1.04 PFlop/s | Cedar (SFU) |
| Resnick HPC Cluster | California Institute of Technology | NVIDIA L40S | 16 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Hortense | Flemish Supercomputer Centre | NVIDIA A100 SXM4 40GB | 80 | 745.57 TFlop/s to 23.60 PFlop/s | Narval |
| Hortense | Flemish Supercomputer Centre | NVIDIA A100 SXM4 80GB | 80 | 1.16 PFlop/s to 1.50 PFlop/s | Champollion |
| Barbora | IT4Innovations | NVIDIA Tesla V100 | 32 | 93.33 TFlop/s to 1.57 PFlop/s | JADE2 |
| bwUniCluster 3.0 | Karlsruhe Institute of Technology | AMD Instinct MI300A | 4 | 123.33 TFlop/s to 321.88 TFlop/s | Cosmos (SDSC) |
| Magerit-3 | CeSViMa | NVIDIA L40S | 20 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Pryor North TPU Campus | Alphabet | Google TPU v5p | 23,500 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Columbus TPU Campus | Alphabet | Google TPU v5p | 28,200 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Columbus TPU Campus | Alphabet | Google TPU v6e (Trillium) | 204,600 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Columbus TPU Campus | Alphabet | Google TPU7x (Ironwood) | 38,900 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google New Albany TPU Campus | Alphabet | Google TPU v5p | 66,600 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google New Albany TPU Campus | Alphabet | Google TPU v6e (Trillium) | 262,700 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Council Bluffs East TPU Campus | Alphabet | Google TPU v6e (Trillium) | 125,200 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Council Bluffs East TPU Campus | Alphabet | Google TPU v5p | 54,300 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Google Council Bluffs East TPU Campus | Alphabet | Google TPU7x (Ironwood) | 41,900 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| AWS Madison Mega Site (Canton) | Amazon | AWS Trainium2 | 326,600 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| AWS Ridgeland Campus | Amazon | AWS Trainium2 | 261,200 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Microsoft Project Osmium | Microsoft | NVIDIA B200 SXM | 45,500 | 1.56 EFlop/s to 7.50 EFlop/s | Betty |
| Microsoft Goodyear AI Datacenter | Microsoft | NVIDIA B200 SXM | 58,100 | 1.99 EFlop/s to 9.57 EFlop/s | Betty |
| Meta Rosemount (Minnesota) | Meta Platforms | NVIDIA B200 SXM | 85,400 | 2.92 EFlop/s to 14.07 EFlop/s | Betty |
| Technion Zeus | Technion | NVIDIA A100 (in DGX A100) | 72 | 1.09 PFlop/s to 10.54 PFlop/s | Cambridge-1 |
| Technion Zeus | Technion | NVIDIA Tesla V100 | 8 | 23.33 TFlop/s to 392.22 TFlop/s | JADE2 |
| Jairu | USP CIAAM | NVIDIA B200 SXM | 96 | 3.29 PFlop/s to 15.82 PFlop/s | Betty |
| Nscale Verne Iceland Cluster | Nscale | NVIDIA GB300 NVL72 | 4,600 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Casper | NCAR | NVIDIA Tesla V100 | 64 | 186.67 TFlop/s to 3.14 PFlop/s | JADE2 |
| Neo | Lawrence Livermore National Laboratory | NVIDIA H200 SXM | 32 | 1.26 PFlop/s to 4.52 PFlop/s | GMO GPU Cloud |
| AICR | MGHPCC | NVIDIA B200 SXM | 248 | 8.49 PFlop/s to 40.87 PFlop/s | Betty |
| Bouchet | Yale University | NVIDIA B200 SXM | 56 | 1.92 PFlop/s to 9.23 PFlop/s | Betty |
| Hopper (Yale) | Yale University | NVIDIA B200 SXM | 48 | 1.64 PFlop/s to 7.91 PFlop/s | Betty |
| Insomnia (Columbia) | Columbia University | NVIDIA L40S | 18 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Skipjack (JHU) | Johns Hopkins University | NVIDIA L40S | 64 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Skipjack (JHU) | Johns Hopkins University | NVIDIA H200 SXM | 24 | 942.19 TFlop/s to 3.39 PFlop/s | GMO GPU Cloud |
| Skipjack (JHU) | Johns Hopkins University | NVIDIA B200 SXM | 128 | 4.38 PFlop/s to 21.09 PFlop/s | Betty |
| Skipjack (JHU) | Johns Hopkins University | NVIDIA B300 SXM | 8 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Quest (Northwestern) | Northwestern University | NVIDIA H100 SXM5 | 96 | 1.95 PFlop/s to 4.44 PFlop/s | Israel-1 |
| Quest (Northwestern) | Northwestern University | NVIDIA A100 SXM4 80GB | 72 | 1.04 PFlop/s to 1.35 PFlop/s | Champollion |
| CSF3 | University of Manchester | NVIDIA A100 SXM4 80GB | 76 | 1.10 PFlop/s to 1.43 PFlop/s | Champollion |
| CSF3 | University of Manchester | NVIDIA A100 SXM4 40GB | 8 | 74.56 TFlop/s to 2.36 PFlop/s | Narval |
| CSF3 | University of Manchester | NVIDIA L40S | 88 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| CSF3 | University of Manchester | NVIDIA H200 SXM | 32 | 1.26 PFlop/s to 4.52 PFlop/s | GMO GPU Cloud |
| Aire (Leeds) | University of Leeds | NVIDIA L40S | 84 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| GenomeDK | Aarhus University | NVIDIA L40S | 16 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Hawk (Cardiff) | Cardiff University | NVIDIA Tesla P100 | 26 | 59.66 TFlop/s to 134.59 TFlop/s | Cedar (SFU) |
| Hawk (Cardiff) | Cardiff University | NVIDIA Tesla V100 | 30 | 87.50 TFlop/s to 1.47 PFlop/s | JADE2 |
| Sunbird | Swansea University | NVIDIA Tesla V100 | 8 | 23.33 TFlop/s to 392.22 TFlop/s | JADE2 |
| Shenzhen 10,000-Card Ascend Cluster | Shenzhen Municipal Government | Huawei Ascend 910C | 10,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Guacolda | University of Chile | NVIDIA Tesla V100 | 4 | 11.67 TFlop/s to 196.11 TFlop/s | JADE2 |
| Nscale Texas GB300 Campus (Microsoft) | Nscale | NVIDIA GB300 NVL72 | 104,000 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Voltage Park H100 Cloud | Voltage Park | NVIDIA H100 SXM5 | 24,000 | 486.3 PFlop/s to 1.11 EFlop/s | Israel-1 |
| Sirius (Tsukuba) | University of Tsukuba | AMD Instinct MI300A | 96 | 2.96 PFlop/s to 7.72 PFlop/s | Cosmos (SDSC) |
| OCTOPUS | D3 Center, Osaka University | NVIDIA Tesla P100 | 148 | 339.59 TFlop/s to 766.10 TFlop/s | Cedar (SFU) |
| Raad2 | Hamad Bin Khalifa University | NVIDIA Tesla V100 | 8 | 23.33 TFlop/s to 392.22 TFlop/s | JADE2 |
| Kebnekaise | HPC2N | NVIDIA Tesla V100 | 10 | 29.17 TFlop/s to 490.28 TFlop/s | JADE2 |
| Genius (KU Leuven) | KU Leuven | NVIDIA Tesla P100 | 24 | 55.07 TFlop/s to 124.23 TFlop/s | Cedar (SFU) |
| Fran | ARL DSRC | NVIDIA L40S | 24 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Ruth | ARL DSRC | AMD Instinct MI250X | 128 | 4.08 PFlop/s to 4.60 PFlop/s | LUMI |
| Barfoot | ERDC DSRC | NVIDIA L40S | 12 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
| Makau | MHPCC DSRC | NVIDIA H100 SXM5 | 64 | 1.30 PFlop/s to 2.96 PFlop/s | Israel-1 |
| Hábrók | University of Groningen | NVIDIA Tesla V100 | 24 | 70.00 TFlop/s to 1.18 PFlop/s | JADE2 |
| Monash M3 | Monash University | NVIDIA Tesla V100 | 56 | 163.33 TFlop/s to 2.75 PFlop/s | JADE2 |
| Niflheim | Technical University of Denmark | NVIDIA H200 SXM | 8 | 314.06 TFlop/s to 1.13 PFlop/s | GMO GPU Cloud |
| Niflheim | Technical University of Denmark | NVIDIA A100 SXM4 40GB | 16 | 149.11 TFlop/s to 4.72 PFlop/s | Narval |
| Niflheim | Technical University of Denmark | AMD Instinct MI210 | 2 | not published | No measured HPL rate exists for this part in this dataset, and no stated proxy applies. Nothing is published. |
The read
The headline number is large and it is also the least interesting thing on this page. 100 machines from 88 operators, and the result lands at around half of every ranked system in this dataset combined. These were disclosed because the operators wanted credit for open hardware, for scale, or for a training run, not because anyone required it. There is no reason to think they are unusual, and every reason to think the ones nobody blogged about are larger.
The more useful finding is the shape of the disclosure. Of the 495 unranked systems here, 351 disclose no hardware at all. Not a performance figure, not an accelerator count, nothing that could be turned into a number by any method. That is the normal case. The disclosing clusters are in this analysis precisely because they are the exception, which means any shadow estimate anyone publishes, including this one, is built on the small minority of operators who chose to say something.
Two ways this number misleads, in opposite directions. It overstates, because these clusters would never actually achieve it: HPL needs tuning, a tightly coupled fabric and a reason to run it, and a training cluster has none of the three. It understates far more severely, because FP64 is not what the hardware was bought for. The same H100s deliver an order of magnitude more arithmetic in BF16 and more again in FP8, which is the precision that training actually uses. Read as "how much of a ranked machine is this hardware equivalent to", the figure is fair. Read as "how much compute is out there", it is far too low.
That is the deeper problem with the whole exercise, and it is not solvable by better estimation. The benchmark that defines the visible world measures a precision the invisible world does not use. Even a perfect shadow list would be denominated in the wrong unit. Why FP64 and FP8 diverged is worth reading alongside this page.
The measured-rate chart is worth a second look on its own terms, independent of the shadow argument. It is a clean generational curve built from nothing but sourced fields on system rows, and it shows something the marketing numbers do not: a factor of roughly forty in real delivered HPL throughput per accelerator between 2010 and 2024, and a visible spread between two systems running identical silicon.