Parts/CPUs/NVIDIA Grace
CPUs · Grace · 2023
NVIDIA Grace
72-core Neoverse V2 Arm CPU with LPDDR5X. Ships mainly as the CPU half of GH200 and GB200.
Why this part
NVIDIA's answer to a problem it created. Once the GPU became the machine, the host CPU's remaining job was to move data to it, and every generation of PCIe made that link the bottleneck. Grace exists so NVIDIA can put its own CPU on the same board as its own GPU with a 900 GB/s coherent link between them, which no third-party CPU can offer. Choosing LPDDR5X rather than DDR5 is the other half: lower power and higher bandwidth per watt, at the cost of being soldered down and non-upgradable. A system that buys Grace has bought NVIDIA's CPU, GPU, interconnect and memory decision in a single purchase order.
Systems using this part
- Systems
- 32
- Rmax underneath
- 2.49 EFlop/s
- Units recorded
- 13,830
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| HoreKa 2 2026-08 | CPU | Germany | 150 | - | NVIDIA | - |
| ROQUO 2026-06-19 | CPU | Japan | 270 | 19.80 PFlop/s | NVIDIA | - |
| Arrhenius 2026-06 | CPU | Sweden | - | 66.82 PFlop/s | NVIDIA | #42 → |
| KDDI GPU Cloud 2026-01 | CPU | Japan | - | 22.21 PFlop/s | NVIDIA | - |
| Azure NDv6 GB300 Cluster 2025-10 | CPU | USA | - | - | NVIDIA | - |
| JUPITER 2025-09 | CPU | Germany | - | 1.00 EFlop/s | NVIDIA | #5 → |
| Isambard-AI Phase 2 2025-06 | CPU | United Kingdom | - | 216.5 PFlop/s | NVIDIA | #13 → |
| Discovery 6 2025-06 | CPU | USA | - | 164.2 PFlop/s | NVIDIA | #17 → |
| Miyabi 2025-01 | CPU | Japan | - | 46.80 PFlop/s | NVIDIA | #28 → |
| Isambard 3 2025-01 | CPU | United Kingdom | 384 | - | NVIDIA | - |
| Nano 4 GB200 2025 | CPU | Taiwan | - | 5.16 PFlop/s | NVIDIA | - |
| ROMEO-2025 2024-11 | CPU | France | - | 9.86 PFlop/s | NVIDIA | #2 → |
| JETI 2024-11 | CPU | Germany | - | 83.14 PFlop/s | NVIDIA | #6 → |
| Vista (TACC) 2024-09 | CPU | USA | - | - | NVIDIA | - |
| Isambard-AI 2024-07 | CPU | United Kingdom | - | 7.42 PFlop/s | NVIDIA | #238 → |
| CEA-HE 2024-06 | CPU | France | - | 90.79 PFlop/s | NVIDIA | #29 → |
| Venado 2024-04 | CPU | USA | - | 98.51 PFlop/s | NVIDIA | #26 → |
| JEDI 2024-04 | CPU | Germany | - | 4.50 PFlop/s | NVIDIA | #1 → |
| Alps 2024-03 | CPU | Switzerland | - | 434.9 PFlop/s | NVIDIA | #14 → |
| NASA Cabeus 2023-12 | CPU | USA | 350 | - | NVIDIA | - |
| Shaheen III - GPU | CPU | Saudi Arabia | - | 122.8 PFlop/s | NVIDIA | #22 → |
| DAEDALUS | CPU | Greece | - | 85.69 PFlop/s | NVIDIA | #31 → |
| DeltaAI | CPU | USA | - | - | NVIDIA | #43 → |
| Hon Hai Kaohsiung Super Computing Center | CPU | Taiwan | 2,304 | - | NVIDIA | - |
| HammerHAI | CPU | Germany | - | - | NVIDIA | - |
| Project Ceiba | CPU | USA | 10,368 | - | NVIDIA | - |
| KAIROS | CPU | France | - | 3.05 PFlop/s | NVIDIA | #1 → |
| Triton | CPU | Finland | 4 | - | NVIDIA | - |
| MeluXina-AI | CPU | Luxembourg | - | - | NVIDIA | - |
| Castner | CPU | USA | - | 2.92 PFlop/s | NVIDIA | #436 → |
| Tyche-hsg | CPU | USA | - | - | NVIDIA | - |
| Theia-cmh | CPU | USA | - | - | 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.