NVIDIA Rubin AI Infrastructure Liquid Cooling Innovation Replaces Fans to Improve Data Center Power Efficiency and Sustainability
NVIDIA has radically changed the thermodynamics for all data centers with an innovation in their new use of Rubin artificial intelligence infrastructure having fully liquid cooled. Using a NVIDIA DSX AI factory design, the entire new platform is fan free as it runs the cooling fluid through the inner work at around 45 degrees Celsius (113 Fahrenheit). Knowing this higher temperature point greatly improves the power efficiency of the data center because much less mechanical energy is needed to shunt the hotter heat.
The cooling liquid 75 percent water and 25 percent propylene glycol flows into the server rack at 45 oC (113 oF) and leave at approximately 55 oC (131 oF) having been warmed by the silicon chip. Since liquid cooled cold plates keep the library chip sets within the validated operating region there is no impact to performance even at these more elevated operating temperatures. Ali Heydari, the director of data center cooling and infrastructure at NVIDIA shares the environmental significance
The NVIDIA DSX reference design for AI factories has zero water consumption, we have eliminated massive amounts of power usage and pretty much all water usage. With dry cooler based designs, it is a closed loop system with no evaporative water cooling, outside of maybe 1% of the year when we might need chillers in some climates.
Historically in a data center up to 40% of the total electricity has been used for cooling infrastructure, making it a significant target for operational cost savings. Data from the industry shows that a net increase of just 1 degree in the cooling temperature of the chiller plant reduces costs of cooling energy by about 4%. In a 50 mw hyperscale facility, this equates to more than $4,000,000 annual savings in utilities by implementing such a liquid cooled solution. Additionally, in moderate climates, the 45 degree architecture allows for water free (chillerless) operations using dry coolers, saving at least 2,600,000 gallons per mw per year while bringing water requirements down to virtually zero.
This trend this thermodynamic shift is contributing to a larger trend sweeping the data center industry. Infrastructure builders are tailoring their catalogues to the power densities demanded by the new silicon. Motivair president and CEO Richard Whitmore an advanced cooling unit of experts within the Schneider Electric group explained that the company had been working with NVIDIA for almost a decade in preparation for this change. As he explained
Once the watts per chip crossed a certain level, liquid cooling became mandatory. In the right geographic location, with the right system design, you do not need any refrigeration equipment. You can just put big radiator coils outside and use the air temperature for all your cooling. It is incredibly efficient.
Removing high velocity cooling fans eliminates many of the serious acoustic issues experienced in traditional facilities. Air cooled data centers regularly reach noise levels of 85 dB or more forcing the use of protective ear gear for the operator. The fanless Rubin architecture runs essentially quietly.
In addition this also allows a dramatic leap forward in compute density. The physical size of water cooling modules is far smaller than the cumbersome air heat sinks, so servers that previously occupied 6 racks of room can be shrunk down to just 2 racks of space.
The engineering team at NVIDIA overcame earlier hybrid cooling issues by creating a closed, clean front panel and a more efficient routing system that supplies fluid to multiple high power chips with just one inlet and outlet loop. Furthermore, the exhaust fluid, being at such a high temperature, has numerous other uses for waste heat recovery. The 55 degree C water leaving the server racks can be conveniently used for direct heating for adjacent commercial offices, municipal building housing residential dwellings, etc.
