Cooling architectures shift toward liquid and hybrid designs as AI rack densities rise
Редакция ЦОД:Главное
An industry technology overview identifies liquid and hybrid cooling as the response to rising thermal densities from HPC and AI workloads. It describes Direct Liquid Cooling for high-density zones alongside conventional air cooling elsewhere in the data center.
- The material is a technology overview, not evidence of a specific data-center build, deployment or equipment order.
- Conventional air cooling is associated in the source with typical IT rack densities of 10–15 kW per rack.
- The source places NVIDIA Blackwell rack density near 100 kW, with roadmaps toward 132 kW per rack.
- Hybrid designs pair Direct Liquid Cooling in high-density zones with air cooling for other infrastructure.
- Active RDHx units are identified as a rack-level option for heat removal.
The growth of HPC, AI, advanced analytics and large-scale modelling is changing data-center thermal design requirements, according to an industry technology overview. The publication does not report a specific deployment, equipment order, construction project or commissioned capacity; it is a sector-level assessment of cooling architectures.
The source contrasts conventional air cooling—described as suitable for typical IT rack densities of 10–15 kW per rack—with next-generation NVIDIA Blackwell platforms, which it says approach 100 kW per rack. It also cites technology roadmaps extending toward 132 kW per rack. At these densities, heat removal becomes a central constraint on the deployment of GPU compute infrastructure.
The architecture described is hybrid rather than a wholesale replacement of air cooling. Direct Liquid Cooling is deployed in high-density areas, while air cooling remains in place for the balance of the facility. This approach could allow operators to concentrate liquid-cooling infrastructure around dense GPU clusters while continuing to use air-based systems for lower-density IT loads.
The overview also identifies active rear-door heat exchangers (RDHx) as a rack-level heat-rejection technology. Together with liquid-cooling loops and thermal-management controls, such equipment could see increased relevance as operators align cooling and power-system design around higher-density AI and HPC deployments.
The source states that liquid is more than 25 times more effective at heat transfer, but provides no methodology, comparison conditions or defined heat-transfer metric for that claim. The available material also ends as it begins to address power and cooling architectures, limiting the conclusions that can be drawn from the review.
Thermal management is becoming a deployment constraint for dense GPU infrastructure rather than a secondary operations function. If AI and HPC densities continue to rise, operators may increasingly segment facilities by density, using liquid cooling for high-density clusters while retaining air cooling for less demanding equipment.
- — / 24 сентября 2026 г. · АнглийскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.