If I told you that the solution to cooling the most powerful processors on the planet is to use hot water at jacuzzi temperatures , you'd probably think I'd gone mad. However, that's precisely the premise Nvidia is putting forward to solve one of the biggest headaches of the artificial intelligence era: extreme heat and the exorbitant consumption of water resources.
Generative AI and massive language models have driven server racks to power densities that render conventional air conditioning completely obsolete. We're talking about systems that generate so much heat that direct-to-chip liquid cooling is no longer a luxury or an experiment for enthusiasts, but has become the mandatory standard for anyone wanting to build a competitive infrastructure today.
The concept of the "jacuzzi effect" and applied physics
Nvidia's move is simply brilliant because of how counterintuitive it is. While the liquid in a gaming PC typically hovers around 25 or 30°C, Jensen Huang's company's new AI servers operate with an inlet fluid at 45°C . How is it possible for something so hot to cool a chip? The key lies in the thermal difference: AI processors get so hot that a liquid composed of 75% water and 25% propylene glycol is able to absorb that heat effortlessly, exiting the circuit at around 55°C without affecting the silicon's performance.
Starting from such a high base temperature, the system can take advantage of the difference with the outside temperature to operate passively in most temperate climates. This allows for the replacement of noisy, energy-intensive fans, which typically generate noise levels above 85 decibels, with giant external radiators known as dry coolers , thus optimizing the work environment and reducing electricity bills.
Rubin Architecture: Goodbye to fans
With the arrival of the Rubin architecture, Nvidia has made a bold statement by introducing the first infrastructure with 100% liquid cooling . Unlike previous hybrid systems, where only the GPUs or CPUs were cooled and the rest relied on air, here absolutely all components and network chips are integrated into the circuit. To achieve this, they have sealed the server motherboards and redesigned the liquid flow path to have a single inlet and outlet, resulting in clean, sealed front panels that allow for significantly more processing power to be packed into the same space.
Precision Engineering: CDUs and Cold Plates
To prevent this deployment from ending in a water disaster, meticulous coordination is essential. At the heart of the system are the Refrigerant Distribution Units (CDUs) , which act as the brain, managing pressure, flow, and fluid treatment, and capable of handling power levels from 105 kW to 2,3 MW. Furthermore, cold plates with microstructures are used , capable of capturing up to 90% of the heat directly from the processor.
- Rack manifolds: Stainless steel collectors that prevent leaks and distribute fluid to each GPU.
- Direct-to-Chip Systems: Technology that turns out to be 3.000 times more efficient than air at moving thermal energy.
- RDHx (Rear Door Heat Exchangers): Heat exchangers located at the rear door that remove heat with capacities up to 75 kW.
This entire deployment has a direct economic impact. By no longer wasting energy cooling the room, that electricity budget can be redirected to powering more GPUs, which can result in a 33% increase in computing output per network connection.
The battle against water consumption and the WUE
One of the most critical issues is the environmental impact. Training models like GPT-3 evaporated up to 700.000 liters of fresh water. To measure this, the industry uses Water Usage Effectiveness (WUE) , which calculates the liters consumed per kWh of IT energy. While traditional evaporative cooling has a very high WUE, Nvidia and Microsoft's closed-loop systems aim to approach the ideal value of 0.0 L/kWh.
Essentially, these systems only require water during the initial plant fill. By continuously recirculating the coolant, evaporation towers are eliminated, allowing for a near 100% reduction in water consumption within the data center. Giants like Google and Meta have already committed to achieving water positivity by 2030, aiming to return more water to nature than they consume.
The energy challenge and radical alternatives
It's not all sunshine and roses. Even if the water problem is solved within the four walls of the data center, energy consumption remains a difficult beast to tame. Generating electricity using natural gas or coal still requires enormous quantities of water (up to 2,2 liters per kWh in coal-fired plants), which means that the overall water footprint shifts towards energy generation.
To combat this, even more aggressive alternatives are emerging, such as immersion cooling . In this case, servers are completely submerged in dielectric fluids that do not damage the electronics. Whether through single-phase or two-phase immersion (where the fluid boils at very low temperatures), the goal is to completely eliminate the need for water and reduce energy consumption by up to 82% in certain specific cases.
The transformation of data centers into veritable AI factories is redefining the profitability of technology infrastructure. By optimizing heat flow from silicon to the outside and raising operating temperatures to reduce mechanical stress, the industry is enabling computing power to scale without water and energy consumption hindering the advancement of language models.
