Reducing computing’s carbon footprint isn’t all about optimizing hardware and software. 

When it comes to reducing carbon emissions, tech companies have started considering their complete carbon footprint. Since companies have stronger operational control over their own facilities and energy procurement, many of them have spent the last decade focusing on reducing their emissions related to operational energy consumption (opex) and setting carbon neutral or net zero operational goals. But as more companies are approaching their 100 percent renewable energy targets, they’ve started looking into emissions related to their value chain or capital energy consumption (capex) — indirect emissions that come from hardware manufacturing and infrastructure. 

Researchers from Facebook, Harvard University, and Arizona State University (ASU) have demonstrated that unless both opex and capex emissions are addressed, the tech industry’s carbon footprint will continue to grow. The solution, they say, is to examine ways of reducing carbon emissions that go deeper into the manufacturing supply chain. 

As computer hardware and software become more powerful, they also increase their energy demand. This is particularly true when it comes to the hardware that runs advanced AI and machine learning applications and trains deep learning models. This means engineers and researchers spend a lot of time working on ways to help systems operate with as much energy efficiency as possible. This, coupled with the use of renewable energy, can have a significant impact on opex emissions — those that come from recurring operations.

Data centers like Facebook’s, for example, have increased their energy efficiency through a combination of system improvements and renewable energy. Using warehouse-scale systems and lowering cooling and facility overheads makes for less power consumption. And using renewable energy further reduces a data center’s carbon footprint. In 2019, most of Facebook’s data centers reached nearly zero carbon emissions after shifting to green, renewable energies like solar and wind.

This is a milestone achievement, but it also points the road to the next challenge for tech companies: shifting focus towards capex emissions and setting ambitious net zero targets through their value chain.

What we talk about when we talk about emissions

Data centers provide an easy way to understand the distinction between opex and capex emissions. The energy consumed by hardware inside the data center makes up its opex emissions. But getting all that hardware built and installed, as well as building the data center itself, contributes to greenhouse gas emissions as well. These are the capex emissions.

The Greenhouse Gas (GHG) Protocol, a global standard for measuring and reporting greenhouse gas emissions, categorizes emissions into three scopes. Scope 1 emissions are a company’s direct emissions, such as fuel and chemicals (e.g., gas used by vehicles, refrigerants used to cool offices and data centers, and chemicals used to manufacture semiconductors for semiconductor companies). Scope 2 emissions are indirect emissions from purchased energy and heat that drive semiconductor manufacturing, offices, and data center operations. Scope 3 accounts for all other indirect emissions, including those that come from the supply chain, and those associated with employee business travel, commuting, logistics, purchased goods and services, and capital goods.

A comparison of Scope 1, 2, and 3 emissions for chip companies, mobile vendors, and data center operators, according to the GHG protocol.

In alignment with the GHG Protocol, conducting life cycle assessments (LCAs) is one way to examine a hardware system’s total carbon emissions across its life cycle, including its production and manufacturing, transport, use, and end-of-life processing, or the construction of a data center. LCAs can provide a detailed understanding of the areas and components that contribute most to carbon emissions. 

The shift toward capex emissions

Researchers looked at publicly available sustainability reports and life cycle analyses from AMD, Apple, Facebook, Google, Huawei, Intel, Microsoft, and TSMC. Their meta-analysis found that, for many use cases across the edge and cloud computing spectrum, most carbon emissions came from hardware manufacturing (capex), not operational system use (opex).

Personal devices like smartphones, desktops, and laptops contribute most of their carbon footprint through their manufacturing and use. While emissions from always-connected devices come mainly from opex consumption, emissions from battery-operated devices come mainly from manufacturing (capex). This hardware manufacturing footprint increases as devices become more powerful (having more memory, bandwidth, and/or storage).

While companies have been optimizing their devices’ hardware and software to maximize performance, they also need to focus on the increasing percentage of emissions that come from hardware manufacturing. The researchers estimate that, given the energy efficiency improvements from software and hardware innovation in the last decade, mobile devices, for example, would have to be used three years beyond their typical lifetime to amortize the carbon footprint created by their manufacture.

Data centers have followed a similar trend. The positive impact of renewable energy has shifted the focus on their carbon footprint almost entirely to a need to reduce capex emissions. The construction of the data center itself and manufacture of the hardware that goes into it are responsible for the majority of the data center’s carbon footprint.

Renewable energy has also had a significant impact in the hardware manufacturing sector, where semiconductor factories, for example, have shifted to renewable energy. But the meta-analysis reveals that even under optimistic projections, hardware manufacturing will still account for a large portion of hardware life cycle carbon footprints.

We need optimization at all levels

So, what can be done on the capex end? Facebook’s, Harvard’s, and ASU’s researchers suggest that it will require further work into making hardware more efficient, flexible, and scalable, from their design and manufacturing up to the software level – across the entire computing system stack. 

When looking at semiconductor and other hardware manufacturing, the researchers recommend that hardware needs to be designed from the start with reducing capex emissions in mind. In addition to operational computation performance, data center buildings and hardware supply chains need to be designed with both high performance and low carbon emissions in mind. Using building and infrastructure materials with lower carbon impacts, building repairability and recyclability principles into design processes, extending the life span of hardware, and ensuring responsible end-of-life management will all be essential.

For software, optimizations to algorithms and applications that run data centers and improvements to runtime systems like schedulers, load balancing services, and operating systems can all improve both opex- and capex-related carbon footprints. The researchers also point to recent work into developing new programming languages to allow programmers to write more energy-efficient code.

Reducing the carbon footprint is only the first step

This isn’t a problem that any one company or entity can solve. It will take an industry-wide effort to reduce capex emissions and to continue to reduce opex emissions. There needs to be broader participation, standardization, and disclosures to help the tech industry continue to improve its sustainability performance.

Carbon footprint isn’t the only environmental concern when it comes to making computing more sustainable. Addressing computing’s broader environmental impact, such as air and water pollution, as well as the consumption of limited natural resources, like aluminum, cobalt, copper, glass, gold, tin, lithium, zinc, and water, is also very important.  

The carbon footprint needs to be a first-class design metric, be comprehensive, and take into account not only the performance of hardware, but also the factories where the hardware is built and the data centers where the hardware is deployed.