Parts/Accelerators/NVIDIA H200 SXM
Accelerators · Hopper · 2024
NVIDIA H200 SXM
141 GB HBM3E refresh of the H100, pin-compatible with it. Memory capacity rather than compute was the binding constraint it addressed.
Why this part
The same compute die as H100 with more and faster memory, which tells you where the bottleneck actually was. Inference on a large model is memory-bandwidth bound rather than arithmetic bound, so nearly doubling capacity and adding 40 percent bandwidth improves real throughput more than a new compute architecture would have.
Systems using this part
- Systems
- 42
- Rmax underneath
- 1.04 EFlop/s
- Units recorded
- 18,876
22 of 42 edges disclose a quantity
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| ASPIRE 2B (GPU Partition) 2026-06 | Accelerator | Singapore | - | 66.40 PFlop/s | NVIDIA | #44 → |
| Cassava AI Factory 2026-06 | Accelerator | South Africa | - | 77.79 PFlop/s | NVIDIA | #36 → |
| Maximus-02 2026 | Accelerator | USA | - | 156.1 PFlop/s | NVIDIA | #18 → |
| Neo 2026 | Accelerator | USA | 32 | - | NVIDIA | - |
| Nano 4 2025-11 | Accelerator | Taiwan | - | 81.55 PFlop/s | NVIDIA | #33 → |
| PERUN 2025-11 | Accelerator | Slovakia | 208 | 10.21 PFlop/s | NVIDIA | #162 → |
| TELUS Sovereign AI Factory (Rimouski) 2025-09-24 | Accelerator | Canada | - | 22.74 PFlop/s | NVIDIA | #78 → |
| Cannon2-H200 2025-08 | Accelerator | USA | 96 | 4.24 PFlop/s | NVIDIA | #302 → |
| Alem.Cloud 2025-07 | Accelerator | Kazakhstan | - | - | NVIDIA | - |
| ABCI 3.0 2025-06 | Accelerator | Japan | - | 145.1 PFlop/s | NVIDIA | #19 → |
| Nebius Kansas City Cluster 2025-04 | Accelerator | USA | - | - | NVIDIA | - |
| FPT AI Factory Japan 2025 | Accelerator | Japan | - | 49.85 PFlop/s | NVIDIA | #50 → |
| ISEG2 2025 | Accelerator | Netherlands | - | 202.4 PFlop/s | NVIDIA | #15 → |
| davinci-2 2025 | Accelerator | Italy | - | 14.21 PFlop/s | NVIDIA | #123 → |
| Torch 2025 | Accelerator | USA | 272 | 10.79 PFlop/s | NVIDIA | #133 → |
| Tillicum 2025 | Accelerator | USA | - | 8.85 PFlop/s | NVIDIA | #185 → |
| AI-Farabium 2025 | Accelerator | Kazakhstan | 400 | 17.93 PFlop/s | NVIDIA | #103 → |
| GMO GPU Cloud 2024-11 | Accelerator | Japan | 256 | 10.05 PFlop/s | NVIDIA | #34 → |
| Tesla Cortex 2024-10 | Accelerator | USA | 16,000 | - | NVIDIA | - |
| ARF-ACC 2024 | Accelerator | Turkey | 192 | 12.09 PFlop/s | NVIDIA | #145 → |
| Helma 2024 | Accelerator | Germany | 384 | 32.22 PFlop/s | NVIDIA | #51 → |
| Reindeer 2024 | Accelerator | USA | - | 45.59 PFlop/s | NVIDIA | #32 → |
| Kempner AI Cluster 2023 | Accelerator | USA | - | 16.29 PFlop/s | NVIDIA | #85 → |
| Delta 2022-11 | Accelerator | USA | 8 | 3.81 PFlop/s | NVIDIA | #141 → |
| Discoverer 2021 | Accelerator | Bulgaria | 32 | 4.52 PFlop/s | NVIDIA | #91 → |
| Berzelius 2021 | Accelerator | Sweden | 128 | 5.25 PFlop/s | NVIDIA | #82 → |
| HoreKa 2021 | Accelerator | Germany | 60 | 8.03 PFlop/s | NVIDIA | #52 → |
| Phoenix (PACE) 2020 | Accelerator | USA | 56 | 1.84 PFlop/s | NVIDIA | #277 → |
| Tinkercliffs 2020 | Accelerator | USA | 48 | - | NVIDIA | - |
| farm22 | Accelerator | United Kingdom | 80 | - | NVIDIA | - |
| AzInTelecom Supercomputer Center | Accelerator | Azerbaijan | - | - | NVIDIA | - |
| Serbia National AI Platform | Accelerator | Serbia | 48 | - | NVIDIA | - |
| Nebius Mäntsälä Cluster | Accelerator | Finland | - | - | NVIDIA | - |
| Triton | Accelerator | Finland | 120 | - | NVIDIA | - |
| Redtail | Accelerator | USA | 264 | 12.85 PFlop/s | NVIDIA | - |
| Katana | Accelerator | Australia | - | - | NVIDIA | - |
| Nano 5 | Accelerator | Taiwan | - | 13.06 PFlop/s | NVIDIA | #135 → |
| Nebius Iceland Cluster | Accelerator | Iceland | - | - | NVIDIA | - |
| Vader | Accelerator | USA | - | 3.14 PFlop/s | NVIDIA | #365 → |
| DAIS | Accelerator | Germany | 136 | 6.03 PFlop/s | NVIDIA | #247 → |
| CSF3 | Accelerator | United Kingdom | 32 | - | NVIDIA | - |
| Skipjack (JHU) | Accelerator | USA | 24 | - | NVIDIA | - |
“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.