After the Blackwell architecture pushed single-chip power consumption to 1,000W to 1,200W, the Rubin architecture (VR200) is expected to further raise the single-chip thermal design power (TDP) to an astonishing level of 2,000W or even 2,300W. Such extreme power density has rendered traditional air cooling completely obsolete for core computing, and even existing single-phase liquid-cooled cold plates are facing performance bottlenecks.
Against this backdrop, the combination of micro-channel cold plates (MLCP) and gold-plated integrated heat spreaders (IHS) has been finalized. This shift marks a transformation in thermal engineering from a back-end "peripheral accessory" to a front-end "chip-level packaging integration".
Emax has successfully developed an 800VDC liquid-cooled load bank that can fully simulate the Rubin server rack.
600kW Rack: The Ultimate Challenge of Power and Water
The Vera Rubin NVL72 rack-level solution integrates 72 GPUs into a single rack, with continuous operating power consumption reaching 600kW or higher. This is equivalent to the total electricity use of nearly 500 average households. To support such massive power consumption, data center infrastructure must undergo the following upgrades:
- 48V DC Power Distribution: Traditional 12V power distribution causes unacceptable line losses at such high currents. Rubin mandates a 48V direct-to-chip power architecture, and even power delivery components themselves require liquid cooling protection.
- Industrial-Grade Fluid Distribution System: Each rack row requires 12-inch diameter piping to deliver 1,000 gallons of coolant per minute. This flow rate approaches that of small industrial cooling towers.
- Structural Load-Bearing & Protection: Due to the significant added weight of liquid and manifolds from the liquid cooling system, floor load capacity must reach 2000kg/m². A secondary containment system must also be implemented to address the risk of potential coolant leaks involving thousands of gallons.
Energy Efficiency in an All-Liquid-Cooled Environment
According to NVIDIA’s data, with a 100% liquid cooling configuration, the Vera Rubin platform achieves an extremely low PUE (Power Usage Effectiveness) of 1.1, without a proportional increase in cooling energy or water consumption. This efficiency gain is primarily attributed to removing heat directly from the source (the chips), eliminating energy consumption by air conditioning systems and large fan walls.
Emax also offers liquid-cooled load banks in various specifications for different types of data centers.
