Jak nowe podejście do oceny cyklu życia wspiera bardziej zrównoważone chłodzenie centrów danych

Data centers are the humming, flickering, real-world nerve centers of the “cloud”. They are where digital data is stored and processed, and where an increasing proportion of computing tasks are carried out. As more aspects of our lives move online, and technologies such as machine learning, artificial intelligence (AI), autonomous vehicles and augmented reality become more integrated with our day-to-day lives, demand is rapidly growing for larger, more powerful facilities. 

I have been researching and working in sustainability for over 20 years, and I’m excited by the potential of digital technology to help me do my job better, and to support new solutions for decarbonization and restoring nature. But the proliferation of data centers also presents a challenge: they typically consume 10-50 times more energy per square foot than a commercial office building. Data centers accounted for approximately 1.5% of global power demand in 2020, and this is expected to increase as the AI revolution continues. 

Reducing this footprint while improving performance has become a central concern for data center operators and designers, which has led to a 20% decrease in their energy intensity globally since 2010. But the technology sector still needs to do more. According to the Science Based Targets Initiative (SBTi), it must cut greenhouse emissions by 42% by 2030 (against a 2015 baseline) to achieve net zero and therefore limit global warming to 1.5°C. In practical terms, that means designing data centers to use less energy and water. 

Cooling is particularly important because server racks generate significant heat that must be removed so that they can function effectively. The combined energy use of cooling and power conditioning can be substantial and use 10–20% as much energy as the IT equipment itself in large-scale, efficient datacenters and up to 80–120% as much electricity as the IT equipment itself in small-scale, less efficient data centers. But without understanding the overall environmental impacts of new technologies, it’s hard to chart the most effective path to reach reduction goals. 

Partnering to understand data center impacts 

Data center cooling design is a complex multi-optimization problem: many factors interact to influence performance and sustainability. To help clarify where the biggest opportunities are to improve these metrics, Microsoft and WSP collaborated on a pioneering life cycle assessment (LCA), recently published in Nature, exploring the environmental impacts of four data center cooling techniques for chips used for general compute, or CPUs, and the data centers that host cloud services. This is one of the outputs of our multi-year global strategic partnership, which includes supporting Microsoft to responsibly deliver mission-critical facilities. 

To our knowledge, this is the first publicly available cradle-to-grave LCA of data center cooling technologies conducted by a hyperscale cloud provider. It quantifies the stated environmental impacts for data centers, including all components that make up the delivery of a Virtual Core, such as the chips, servers, supporting equipment and buildings from embodied to operational, plus transportation and end of life for a data center. It features the ability to examine these environmental impacts under different alternative novel cooling technologies. Our work took several years and goes into much greater detail than previous high-level studies. Microsoft has set the impressive goal of being carbon-negative and water-positive by 2030 – kudos to the team for really sharpening their pencils and doing the due diligence as they continue that journey. 

Decarbonization is an urgent challenge, so we need to understand the levers that will have the most impact, most quickly. LCAs are essential because they allow us to evaluate the whole footprint of a product or system (and not just carbon but a range of environmental impacts), pinpoint the most effective reduction strategies, and avoid shifting impacts from one life cycle stage to another. They’re invaluable at design stage because they enable sustainability to be factored into decisions alongside feasibility and cost analysis. 

Microsoft intends to apply the insights from this research to optimize the cooling design of future data centers, supporting its broader sustainability goals. WSP’s purpose is to shape communities to advance humanity, so I’m really pleased that the paper is also accompanied by a free, open-source, Excel-based tool that any data center operator or designer can use to evaluate different server cooling technologies and locations and consider the trade-offs between them. It can also help purchasers of cloud services or data center capacity to understand the emissions or water implications in the context of their own sustainability targets. WSP works with organizations in many sectors to quantify their direct and indirect impacts, and to develop strategies to mitigate them, and we can help apply this tool to different contexts. 

Headline findings

We looked at three technologies – cold plate, one-phase immersion and two-phase immersion – comparing them against traditional air-cooling. Each has pros and cons, but it has been assumed that their benefits outweigh the drawbacks. We set out to quantify the environmental impacts to inform better decisions. 

Our research found that using liquid cooling in data centers could reduce greenhouse gas emissions by 15-21%, energy demand by 15-20%, and blue water (freshwater) consumption by 31-52%. For a typical hyperscale facility, these reductions could be broadly equivalent to removing up to 4,500 cars from US roads annually, powering around 22,200 US homes, or saving approximately 20 Olympic swimming pools of water. These figures are illustrative and based on specific assumptions outlined in the study.

When we’re producing LCAs for complex systems, one of the first challenges is to identify a single unit for comparing systems that perform similar functions in different ways. Here, we chose the annualized virtual core (Vcore) as our “functional unit”. This is the basic unit of the cloud, representing the amount of computing power used over a year, and it enables comparison of different data center configurations and locations across the entire ecosystem.

The graph below summarizes our findings: it shows the GHG emissions, total energy demand and blue water consumption impacts for each technology, per Vcore per year. 
We modeled two different power sources for the operation of the data center: one with 0% renewable energy as a baseline (representing a fossil-based grid), and one with 100% renewable energy (reflecting Microsoft’s commitment to renewable energy). All of the results are shown relative to the impact of an air-cooled data center powered by 0% renewable energy. 

This is illuminating because it allows us to see the trade-offs in other areas such as water consumption. 

Using water-based evaporative cooling instead of electric cooling systems that use air or refrigerants instead is often presented as the way to solve “thirsty” data centers. But just because there’s no operational water use within the data center, that doesn’t mean there isn’t any water involved in cooling. Data centers also consume water indirectly, via electricity generation. Cooling technologies that use more power could result in higher water consumption overall, depending on the source of that electricity. 

Geographical location is relevant too. We need to look at exactly where that water is being consumed – is it in an area that’s currently experiencing water stress, due to climate, a concentration of industrial users, or both? 

The research found that greenhouse gas emissions during the operational phase of a data center could be reduced by 85-90% by switching to renewable energy, regardless of the cooling technology used. However, embodied emissions from hardware and infrastructure remain significant.

This is why it’s important to stress that we shouldn’t stop just at cooling technology and the energy supplied to power it. First, we need to think about using renewable energy further up the supply chain – we know that the hotspot for servers is the energy it takes to produce them. 

And we still need to optimize data centers – and every other part of the supply chain – to reduce energy consumption, because other industries are scaling up and switching to renewables at the same time. We need massive investment in renewable energy all over the planet and efficiency remains essential so that we can meet growing demand sustainably.

Nuclear energy is emerging as a possible solution for powering data centers, and some data center providers are already scaling with small nuclear reactors. It would be interesting to apply our LCA methodology to compare renewable energy sources versus nuclear. We could also include emerging technologies like heat recovery – WSP is currently working on a project in Manhattan to supply heat from two data centers to a thermal energy network. 

A surprising finding from our research is that the impacts of the building itself are almost non-existent when amortized down to our functional unit, the virtual core. You can see them – but only just – in the tiny dark red bands in the graph above. 

That doesn’t mean that we shouldn’t focus on decarbonizing the built infrastructure. Buildings are responsible for more than a third of global energy consumption and emissions, and the scale of data center development that is planned around the world will add up to a significant impact. WSP’s Sustainable Building Consulting teams are helping clients in many sectors, including data centers, to understand the contribution their buildings make to climate change and develop strategies to mitigate it.

This study shows us that in the overall context of decarbonization for a data center operation, there is a clear hierarchy of mitigations when it comes to choosing what could have the most impact. It reaffirms what we already suspected: that use-phase energy dominates, and that switching to renewable energy is an important step forward in decarbonization. And the study also highlights where we need to look next: the second largest contributors were in the components: servers, chips, printed wiring boards. We should engage with suppliers across the value chain, so that we can better understand how to decarbonize those, too. 

One of the biggest challenges we faced is the limited availability of current environmental data on the electronic equipment used in data centers. 

For example, most of the publicly available data about semiconductor production is outdated and may not reflect current technology. In an ideal world, suppliers would produce full environment product declarations (EPDs) for all of these components, quantifying impacts across many categories, including carbon, energy and water. But carbon footprint information at least would be a start. 

It’s also really important to mitigate the F-gases used in the production of semiconductors, such as sulfur hexafluorine. SF6 is one of the most potent greenhouse gases, with a global warming potential 23,500 times higher than CO2 – which means that it will trap 23,500 times more heat in the atmosphere. What could the end purchasers do to encourage the supply chain to engage on more and better abatement of these F-gases? 

In future LCAs, it could be relevant to include other environmental categories such as smog formation potential. Air quality is a really important factor in siting a data center, because it pulls in air from outside for cooling. If there are a lot of particulates or other pollutants, that can cause the equipment to deteriorate and hinder the functioning of the data center. Or it could lead to increased costs (and energy) to filter the air as it enters. 

This paper is timely because we are at the start of a whole new evolution in data centers, and there is a lot we still need to know. 

In this study, we’ve looked at chips for general compute, or CPUs, and the data centers that host cloud services. We’re now in an era of much more powerful facilities used for AI. Some of our conclusions could be applied to AI, but those data centers and ecosystems will be set up differently, warranting ongoing focused analysis. 

As AI applications ramp up, this new class of data center is using even more energy and water and materials, and they are only going to scale faster and faster. We should make sure this doesn’t happen at the cost of the environment, and that means giving everyone involved in data center design – from architects and engineers to procurement teams, operators, policy markers and more - information about environmental impacts so they can integrate it into all the parameters they’re considering. 

WSP is also working with NVIDIA, the leading manufacturer of specialized chips for AI. Together, we have completed the world’s first cradle-to-gate product carbon footprint of a GPU, for its flagship H100 and B200 chips. 

It’s great to be contributing to the ongoing efforts to improve sustainability in data center cooling design at such a pivotal time for the industry. I’m personally excited that a tool we’ve created is now available for anyone to use to evaluate options and trade-offs for cooling servers. And I’m utterly impressed by the clever and dedicated teams at WSP and Microsoft who came together not only to do this work in a rigorous and transparent way, but also that this achievement was released by Microsoft into the public domain where it can do the most good. I’m looking forward to continuing our pioneering work with Microsoft, and to supporting other organizations to follow its lead. 

As a Senior Vice President and Technical Fellow, Julie brings multiple skills to WSP's sustainability, energy and climate change offerings. Her expertise includes Life Cycle Assessment (LCA) and carbon and water footprinting of products and services, in addition to sustainability strategy and decarbonization in the tech/electronics, bio-based fuels and chemicals, agriculture and food, textiles and apparel, and building and construction sectors. She has been conducting LCAs since 2007. 

Julie holds a Ph.D. in Biological Systems Engineering with a distributed minor in Environmental Analysis and Policy and a M.Sc. in Agroecology from UW-Madison, and a B.Sc. in Natural Resource Management with a Concentration in Conservation and Applied Ecology with a minor in Environmental Policies, Institutions and Behaviors from Cook College of Rutgers University. 

Her research work and publications focus on applying LCA to the cloud, data centers and novel liquid cooling technologies, applying AI to streamline LCA workflows, LCA of novel data storage in synthetic DNA, water embedded in electricity (Scope 2 water), sustainable food systems and alternative proteins, LCA of bio-based fuels, energy policy analysis, and sustainability assessment methodology.

She is also a member of the Scope 3 Technical Working Group for the revision of the GHG Protocol and has been a member of the ISO/TC Technical Advisory Group (TAG) 207 on Environmental Management since 2019.