Parts/Cooling/Direct-to-chip cold plate
Cooling · DLC · 2018
Direct-to-chip cold plate
Cold plate on the package, air for the balance of the node.
Why this part
A partial commitment: liquid on the processors and accelerators, air for everything else. Summit and Sierra are built this way. It is the pragmatic option when the hot components are a minority of the node, and it stops being enough once memory, network adapters and voltage regulators are themselves substantial heat sources, which is roughly what happened by 2022.
Systems using this part
- Systems
- 11
- Rmax underneath
- 273.6 PFlop/s
- Units recorded
- -
0 of 11 edges disclose a quantity
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| IREN Childress Horizons 1-4 (Microsoft) 2026-08 | Cooling | USA | - | - | IREN | - |
| Cerebras Oklahoma City 2025-09 | Cooling | USA | - | - | Cerebras | - |
| TensorWave Tucson MI325X Cluster 2025-07 | Cooling | USA | - | - | TensorWave | - |
| Stargate Abilene 2025-06 | Cooling | USA | - | - | Crusoe | - |
| xAI Colossus 2024-07 | Cooling | USA | - | - | Supermicro | - |
| Karolina 2021-08 | Cooling | Czechia | - | 6.75 PFlop/s | Hewlett Packard Enterprise | #69 → |
| Frontera 2019-06 | Cooling | USA | - | 23.52 PFlop/s | Dell Technologies | #5 → |
| Sierra 2018-06 | Cooling | USA | - | 94.64 PFlop/s | IBM | #62 → |
| Summit 2018-06 | Cooling | USA | - | 148.6 PFlop/s | IBM | #58 → |
| Bluefire 2008-04-24 | Cooling | USA | - | 57.73 TFlop/s | IBM | #44 → |
| Lambda Chicago AI Factory | Cooling | USA | - | - | EdgeConneX | - |
“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.