English
Español
Deutsch
Français
日本語
The rapid expansion of artificial intelligence computing has pushed modern data centers to unprecedented levels of power density. Traditional cooling technologies are no longer sufficient to support next-generation AI infrastructure. With the launch of the NVIDIA Rubin architecture and the NVIDIA Vera Rubin NVL72 platform, NVIDIA has fundamentally redefined how high-performance computing systems are cooled. Instead of treating cooling as an auxiliary system, Rubin integrates liquid cooling as a core part of the computing architecture itself.
The Thermal Challenge of Next-Generation AI Chips
AI workloads such as large language model training, generative AI inference, and robotics simulations require enormous computational power. As a result, GPU power consumption has increased dramatically.
Next-generation AI processors in the Rubin series are expected to reach thermal design power levels exceeding 2,000 watts per GPU, far beyond the capabilities of traditional air cooling systems.
At these power densities:
● Air cooling becomes inefficient and energy-intensive
● Heat flux concentration increases dramatically
● Data center power efficiency declines due to cooling overhead
This“thermal wall” is one of the biggest barriers to scaling AI infrastructure.
To overcome it, NVIDIA has adopted direct liquid cooling as the default thermal management technology in the Rubin platform.
Direct Liquid Cooling at the Core of Rubin
Unlike traditional servers that rely on high-speed fans and chilled air, Rubin systems utilize Direct-to-Chip Liquid Cooling (DLC). In this approach, coolant flows through cold plates that are directly attached to heat-generating components such as GPUs, CPUs, and networking chips.
The Rubin platform uses warm-water, single-phase liquid cooling, operating with inlet temperatures of around 45°C.
This design provides several advantages:
● Higher heat transfer efficiency compared to air cooling
● Reduced energy consumption for fans and chillers
● Improved rack power density
● More stable performance under sustained workloads
Because the system can operate with warm water, many data centers can eliminate large chiller plants and instead rely on simpler cooling infrastructure.
Rack-Scale Liquid Cooling Architecture
The NVIDIA Vera Rubin NVL72 platform represents a new generation of liquid-cooled AI superclusters.
Each NVL72 rack integrates:
● 72 Rubin GPUs
● 36 Vera CPUs
● NVLink 6 high-speed interconnects
● Advanced networking through BlueField DPUs and ConnectX SuperNICs
These racks deliver multi-exaflop-scale AI performance while maintaining thermal stability through a highly integrated liquid cooling loop.
In many Rubin systems:
● Fans are minimized or eliminated
● Cooling is handled through cold plates, manifolds, and coolant distribution units (CDUs)
● Entire racks operate as liquid-cooled compute modules
This architecture transforms the data center from an air-based cooling environment into a liquid-centric infrastructure.
Microfluidic Cooling: Bringing Liquid Closer to Silicon
To manage extreme heat density, NVIDIA and semiconductor partners are exploring microfluidic cooling technologies for Rubin-class processors.
Instead of cooling the chip from the outside, microchannels are integrated directly into the chip packaging, allowing coolant to flow within microns of the transistors.
These microscopic channels dramatically increase heat removal efficiency by:
● Shortening the thermal path
● Increasing heat exchange surface area
● Reducing thermal resistance between the chip and coolant
This approach marks a significant evolution in thermal engineering for AI hardware.
Efficiency Gains for AI Data Centers
Liquid cooling in Rubin systems provides multiple operational benefits for hyperscale data centers:
1. Higher Performance Stability
Direct liquid cooling prevents thermal throttling during heavy AI training workloads.
2. Lower Energy Consumption
Liquid cooling transfers heat far more efficiently than air, reducing cooling energy requirements.
3. Increased Rack Density
Data centers can deploy more compute power per rack without exceeding thermal limits.
4. Reduced Infrastructure Costs
Warm-water cooling reduces or eliminates expensive chillers and large HVAC systems.
These improvements enable AI infrastructure to scale efficiently as model sizes and computing demand continue to grow.
A New Era of Liquid-First Computing
The Rubin platform signals a fundamental shift in data center design philosophy. Instead of adapting cooling systems to computing hardware, NVIDIA has engineered computing systems that are built around liquid cooling from the start.
As GPU power consumption continues to rise and AI workloads expand, liquid cooling will increasingly become the standard for high-performance computing environments.
For hardware manufacturers, cooling solution providers, and data center operators, the Rubin era marks the beginning of a liquid-first future for AI infrastructure.
+86-152 2030 1231
E-mail
