Virtualization and software-defined whatever aren’t the only ways in which enterprise infrastructure is changing these days. Researchers are getting ever closer to replacing traditional air-cooling methods in the server room with new liquid-cooled designs, promising both greater efficiency and lower costs as hardware footprints scale up and out to meet burgeoning data requirements.
The latest advance comes courtesy of 3M, SGI and Intel, which recently showed off a proof-of-concept approach for distributed memory architectures built around SGI’s ICE X platform. The design allows Xeon E5-2600 boards to be immersed directly in a newly developed dielectric that 3M has dubbed “Novec,” which is said to provide dramatically more effective heat exchange than air and requires less room and is less expensive to operate. To be sure, the system is still in the experimental stage, although it provides a key glimpse into the likely development path of future cloud-scale hardware cooling architectures.
The revolutionary component, naturally, is the dielectric itself. Originally developed as a fire suppressant, Novec enables a highly efficient two-stage cooling process in which heated liquid evaporates off the board, while a separate condenser cools it back into liquid form for return to an awaiting holding tank. 3M claims it can accommodate 100kW/m2 using normal water sources for the condenser, rather than a dedicated supply. The fluid is also said to be non-toxic and leaves no residue on vital computing components.
At the same time, a company called Ebullient is close to offering a commercial liquid-cooling product that is said to boost rack densities and lower operating costs in production server environments. The system uses a specialized refrigerant that is pumped into the rack where it hits a chip-mounted copper plate, absorbing available heat and then shuttling it to an external heat exchanger. As in 3M’s design, Ebullient converts the liquid into vapor and back again as part of the cooling process. However, Ebullient does not utilize a full immersion approach, preferring instead to pump through a series of tubes to heat-generating components.
Other solutions like the Rack DCLC by CoolIT are making their way into high-performance configurations. Penguin Computing’s Relion 2808GT, for example, has seen a 25 percent decrease in operating costs by applying the direct-to-chip liquid solution for high-density Nvidia Tesla K40m boards. This allows Penguin to up the processing power per rack without exceeding heat limits, which in turn enables HPC shops to handle larger loads without adding to their data center footprints. As one of the few authorized solution providers for Facebook’s Open Compute Project, Penguin is likely to be a leading purveyor of scale-out, liquid-cooled infrastructure.
Liquid cooling has a reputation as an exotic practice, more attuned to specialty computing shops rather than the run-of-the-mill enterprise. But as data loads increase and cloud-scale architectures become the norm, it will likely become a mainstream solution in short order.
The next likely step is the integration of liquid solutions into modular infrastructure, which would provide scale-out performance under increasingly tight footprints and with highly favorable operating costs. This won’t look like the data centers we are used to, but in a world where everything is changing, it’s only logical that cooling infrastructure will change as well.