Parts/Accelerators/NVIDIA GH200 Grace Hopper Superchip
Accelerators · Grace Hopper · 2023
NVIDIA GH200 Grace Hopper Superchip
Grace Arm CPU and Hopper GPU on one module over NVLink-C2C, with coherent access to LPDDR5X and HBM3e.
Why this part
A CPU and a GPU sold as one part, linked at 900 GB/s and sharing a coherent address space, so the GPU can reach 480GB of LPDDR5X as if it were its own. It dissolves the distinction between host memory and device memory that has organised heterogeneous computing since Roadrunner. It also dissolves a supply chain: Alps, JUPITER, Venado and Isambard-AI all buy their CPU, GPU, memory and the link between them from one vendor on one purchase order, and there is no configuration of that node with a competitor's part in it.
Systems using this part
- Systems
- 28
- Rmax underneath
- 2.49 EFlop/s
- Units recorded
- 15,299
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| Arrhenius 2026-06 | Accelerator | Sweden | - | 66.82 PFlop/s | NVIDIA | #42 → |
| Roihu 2026-06 | Accelerator | Finland | 528 | - | NVIDIA | - |
| Tigris (RIT) 2026 | Accelerator | USA | - | 5.08 PFlop/s | NVIDIA | - |
| JUPITER 2025-09 | Accelerator | Germany | - | 1.00 EFlop/s | NVIDIA | #5 → |
| Isambard-AI Phase 2 2025-06 | Accelerator | United Kingdom | - | 216.5 PFlop/s | NVIDIA | #13 → |
| Discovery 6 2025-06 | Accelerator | USA | - | 164.2 PFlop/s | NVIDIA | #17 → |
| Olivia 2025-06 | Accelerator | Norway | 448 | 13.20 PFlop/s | NVIDIA | #134 → |
| Miyabi 2025-01 | Accelerator | Japan | 1,120 | 46.80 PFlop/s | NVIDIA | #28 → |
| ROMEO-2025 2024-11 | Accelerator | France | 232 | 9.86 PFlop/s | NVIDIA | #2 → |
| JETI 2024-11 | Accelerator | Germany | - | 83.14 PFlop/s | NVIDIA | #6 → |
| Vista (TACC) 2024-09 | Accelerator | USA | 600 | - | NVIDIA | - |
| Isambard-AI 2024-07 | Accelerator | United Kingdom | - | 7.42 PFlop/s | NVIDIA | #238 → |
| CEA-HE 2024-06 | Accelerator | France | - | 90.79 PFlop/s | NVIDIA | #29 → |
| Venado 2024-04 | Accelerator | USA | - | 98.51 PFlop/s | NVIDIA | #26 → |
| JEDI 2024-04 | Accelerator | Germany | - | 4.50 PFlop/s | NVIDIA | #1 → |
| Helios 2024-04 | Accelerator | Poland | 440 | 19.14 PFlop/s | NVIDIA | #96 → |
| Alps 2024-03 | Accelerator | Switzerland | 10,752 | 434.9 PFlop/s | NVIDIA | #14 → |
| NASA Cabeus 2023-12 | Accelerator | USA | 350 | - | NVIDIA | - |
| Shaheen III - GPU | Accelerator | Saudi Arabia | - | 122.8 PFlop/s | NVIDIA | #22 → |
| DAEDALUS | Accelerator | Greece | - | 85.69 PFlop/s | NVIDIA | #31 → |
| DeltaAI | Accelerator | USA | 608 | - | NVIDIA | #43 → |
| KISTI-6 | Accelerator | South Korea | - | - | NVIDIA | - |
| Santos Dumont | Accelerator | Brazil | 144 | 14.29 PFlop/s | NVIDIA | #52 → |
| farm22 | Accelerator | United Kingdom | 4 | - | NVIDIA | - |
| KAIROS | Accelerator | France | 64 | 3.05 PFlop/s | NVIDIA | #1 → |
| Bede | Accelerator | United Kingdom | 9 | - | NVIDIA | - |
| Katana | Accelerator | Australia | - | - | NVIDIA | - |
| Castner | Accelerator | USA | - | 2.92 PFlop/s | NVIDIA | #436 → |
“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.