Data center cooling – $40M ARPA-E funding list hints at future tech

Nvidia, HP, Flexnode, and JetCool included in 15-project data center cooling development program to dial up energy efficiency gains.
10 May 2023

Green data center vision: developers are exploring innovative ways of reducing data center cooling demands, which currently can account for up to 40% of data center energy usage overall. Image credit: Shutterstock Generate.

Getting your Trinity Audio player ready...

• COOLERCHIPS looking to reduce burden of intensive data center cooling
• ARPA-E announces shortlist of 15 funded projects
• Improved thermal management key to overall success

It’s not often that you get to see the future. But when it comes to data center cooling technology, ARPA-E’s COOLERCHIPS shortlist of 15 funded projects could provide a glimpse of what’s on the horizon. Cooling Operations Optimized for Leaps in Energy, Reliability, and Carbon Hyperefficiency for Information Processing Systems – to give COOLERCHIPS its full title – spells out the ambitions of the ARPA-E program.

Data centers have evolved into an incredibly popular way of structuring the digital age. Cloud solutions free developers from day-to-day infrastructure maintenance and upgrade tasks. Applications can easily access increased data storage and compute, as needed. And information is backed up remotely for security.

At the same time, operators can focus on the infrastructure and benefit from economies of scale and having everything under one roof. Naturally, this includes server architecture, high-performance computing (HPC), and data storage elements. But data center operators also need to be experts in power generation and water treatment, and have systems redundancy built-in to maximize up time.

Managing energy demands

“All electrical energy going into a data center must eventually be rejected as heat to the environment through a cooling system,” writes the COOLERCHIPS team on its project website. “Data center cooling can be energy intensive; it may account for up to 33—40% of overall data center energy usage and consumes hundreds of billions of gallons of fresh water per year.”

The combination of this energy intensity and the global popularity of cloud solutions is raising concerns about the size of the CO2 footprint of data center operations. Progress is being made in supplying facilities with green energy, although there’s still a mismatch between 24/7 data center operation and the intermittent nature of wind and solar power.

But an environmental issue that’s much harder to get away from is water consumption. Data centers generate a huge amount of heat – enough to warm council offices, university buildings, corporate headquarters, and thousands of homes. How do we know? Because all of those examples are actual projects where waste heat from data centers – operated by Amazon, Microsoft, Facebook, and others – contribute to district heating systems.

In the long term, revolutionary processor designs that take inspiration from the human brain could help to make cloud computing less hungry for energy, and run cooler. Recall how the chess grandmaster Gary Kasparov made hard work for IBM’s Deep Blue – all the more impressive when you consider that the human brain consumes the power of a dim lightbulb (20 W) compared with an estimated 15,000 W for Deep Blue. And the latest supercomputers are more power-hungry still.

As the processing power of high-performance chips rises, so will data center power densities. But brain-inspired computing is still in its early stages and is unlikely to be helping commercial operators in the short to medium term. However, there are other approaches – accelerated by development programs such as COOLERCHIPS – that could yield benefits sooner.

For example, infrastructure that enabled higher heat rejection temperatures would reduce the amount of water being used. And the design of interfaces with much lower thermal resistance would help heat to escape much more readily.

Finding ways to improve thermal management will be key to the success of COOLERCHIPS. And avenues being explored include adapting a two-phase immersion cooling system to improve heat spreading (Intel Federal, which has been awarded $1,711,416 through the program). Heat removal is becoming much more targeted too – for example, Nvidia (which received $5 million in funding) is developing a direct-to-chip cooling system that features a two-phase cold plate and involves submerging components in a hermetically sealed sled.

Tech transfer

HP has received $3,250,000 to develop what’s described as ‘an aggressive liquid cooling solution’ that aims to reject server heat to temperatures as high as 40 degC external ambient air. HP’s approach involves the use of microfluidics, leveraging its inkjet expertise, to first couple silicon microchannels to the surface of server devices and – as a future step – embed cooling networks deeper into designs.

And it’s worth noting that while environmental benefits are high on the agenda, advanced data center cooling solutions will help address resilience concerns too. For example, during heat waves the capacity of data centers to reject excess thermal energy is reduced, which can – in worst-case scenarios – mean that operations have to be paused or scaled back.

The COOLERCHIPS program is also looking to combine its thermal management goals with the trend for modular data center designs that are easier to configure and suit a wider variety of locations. For example, micro data center startup Flexnode has been awarded $3.5 million to create prefabricated, modularly designed edge data centers that include a series of novel heat management features.

Flexnode’s developers aim to make use of a manifold microchannel heatsink, chassis-based hybrid immersion cooling, and a cost-effective additive manufacturing-enabled dry cooling heat exchanger system – among other elements. And the setup could be of particular interest to commercial 5G or IoT operators looking to engineer low-latency systems as compact data center modules could be placed nearby thanks to smaller space requirements.