Intel Looks to Immersion Cooling to Improve HPC Efficiency, Density

Intel and immersion cooling specialist GRC have partnered in a deep exploration of the potential uses for immersion cooling technology in High Performance Computing (HPC) environments. The Register reports that the two companies have recently published a joint whitepaper exploring the actual benefits for data centers in adopting immersion cooling technologies instead of much more ubiquitous air cooling. The benefits? Increased power efficiency, lower environmental impact, and higher compute density: a holy trinity by any other name.

Datacenters worldwide already consume an estimated 1.5 to 2% of the world's available power, at a rate of approximately 3400 TW/h. As an aside, much is said about Bitcoin mining's environmental impact. However, it consumes less than 0.12% of that same worldwide pie - even less now that miners have been investing in improving mining efficiency.

HPC's power consumption and environmental impact is central to the argument for improving its energy efficiency - especially as predictions show that HPC's energy footprint could rise towards 13% throughout the next decade. It's easy to see how this could happen, considering the increasing pushes towards Exascale supercomputing and the surfacing and scaling of yet another computational method - quantum computing.

Of course, higher energy efficiency also entails lower operating costs for the same or higher workload performance - and what HPC operator wouldn't want to jump at that possibility? 

Adding fuel to the fire is that around 40% of HPC's energy consumption is devoted not towards actual computing but simply for cooling. 10 to 15% of that value is sourced from the spinning fans in the various elements that constitute a server. This is partly why HPC has been stuck at a Power Usage Effectiveness (PUE) wall of 1.6 for a decade now. And as transistor and chip density increase (think about AMD's MI 300 APUs and Nvidia's Grace Superchips), air cooling is moving closer and closer to its practical limits. There's only so much heat cooling fans can dissipate.

Adding immersion cooling reduces the moving (and thus prone to failure) fans and improves the overall temperatures of the entire system. However, no matter how good airflow design is, there will always be second-tier spots where cool air doesn't reach. It's estimated that air cooling only captures 30% of servers' heat output and is concentrated on the most pressing hotspots.

Losing the moving fans and running dielectric fluid through heatsinks naturally helps reduce dead space between components. This allows for an increase in server densities, which could also lead to lowered infrastructure costs.

Of course, not all is rosy in the immersion cooling land: that's why the tech isn't as widespread as it seems like it could (or maybe should) be. Moises Levy, Omdia's Senior Principal Analyst for Data Center Physical Infrastructure, did caution that immersion cooling does have more specialized installation and maintenance costs. It also requires more specialized surveillance of the cooling mechanism, involving maintenance of the filtering system, software that can detect subtle pressure variations that could indicate a leak, and monitoring the quality of the dielectric (non-electrical conducive) fluids themselves.

As an added bonus, using immersion cooling fluids will also allow datacenters to cut back on their water consumption for power generation and auxiliary cooling - estimated at 5.13 × 108 m3 in 2021.

Intel is in this for the long game, as this latest partnership builds upon the company's early May announcement to invest $700 million in a new, 18,580 square-meter "mega lab" devoted to qualify, test, and demo its datacenter portfolio using a variety of cooling tech. Besides investment in the facility, the company announced that its Taiwan division was developing an open reference design for immersion cooling systems for its chips.