Gartner’s Technology Trends for 2025 – Energy-Efficient Computing – Energy Consumption of Modern Software

This article is the second part of a blog series on energy-efficient computing.

In October 2024, Gartner released its review of the top ten technology trends for 2025. For the first time, the trend report mentioned energy-efficient computing as a trend, ranking it sixth on the list.

The previous part of the series covered the energy consumption of information technology, focusing particularly on the underlying reasons for its growth. This article delves into energy consumption in software.

How Does Software Consume Energy?

Scientific studies typically divide software-related energy consumption into three categories:

• Energy consumed by data centers and cloud services.
  In addition to the electricity used for the actual computation, this category includes the energy required for other data center equipment, cooling systems, monitoring, etc.

• Energy used for data transmission.
  This category covers everything from intercontinental backbone networks to wireless networks in homes and offices, including mobile networks.

• Energy consumed by end-user devices.
  This category includes typical IT devices like computers and mobile phones, as well as smart TVs, gaming consoles, e-readers, and other devices used to access digital services.

Energy Consumption Distribution

Numerous studies have examined the distribution of energy consumption across these categories. The most recent is a 2024 study by Jens Malmodin et al., “ICT Sector Electricity Consumption and Greenhouse Gas Emissions – 2020 Outcome”. Based on 2020 data, the consumption breakdown is as follows:

• Data centers and cloud services: 22%

• Data transmission networks: 24%

• End-user devices: 54%

The emissions from these categories are further divided into operational emissions (from energy use during operation) and embodied emissions (from device manufacturing and logistics). The breakdown is:

• Data centers and cloud services: operational 82%, embodied 18%

• Data transmission networks: operational 82%, embodied 18%

• End-user devices: operational 49%, embodied 51%

It’s evident that the backend systems (data centers, cloud services, and networks) produce significantly higher emissions during use than throughout their lifecycle. In contrast, end-user devices have emissions split roughly equally. Each new device generates half of its lifecycle emissions at the time of manufacturing.

Key takeaway: Slowing down device replacement is an effective way to reduce IT-related greenhouse gas emissions.

Optimizing Energy Efficiency

• Data Centers: The primary source of energy consumption in data centers is computing. Their energy efficiency is measured by the Power Usage Effectiveness (PUE) metric, which indicates the ratio of general overhead energy to computing power. The best hyperscale data centers have achieved a PUE of 1.1, leaving little room for further improvement. Thus, the focus should be on minimizing energy consumed by computing through software optimization and sensible usage.

• Networks: Network energy consumption is largely non-elastic, meaning networks consume the same amount of power regardless of data transmission volume. This stems from network devices’ inability to transition into low-energy standby modes efficiently. The challenge lies in the time required for devices to “wake up,” which is too long relative to the network’s transmission capacity.

Rapid Change in the IT Landscape

The rapid development of the IT sector and the launch of new services quickly alter the landscape. Older estimates, even a few years old, may no longer be relevant. For instance, the Malmodin team’s study predates the explosive growth of AI, meaning current figures might differ. However, in the absence of better data, these estimates remain a solid foundation for analysis.

Micro-Level Energy Consumption – Instructions and Data

On the micro level, energy consumption within individual computers is practically due to the use of the CPU, GPU, memory and storage space:

• 1

  Executing machine language instructions (op codes) by the processor and the GPU. Inefficiently written software increases the number of instructions executed.

• 2

  The amount of data processed by the software. The more data processed, the more instructions are executed and the more information is transferred between device components, such as the processor and memory or the storage.

Efforts to control energy consumption should target both dimensions, steering them toward balance and efficiency.

In the next article, we will explore energy consumption in artificial intelligence and methods for mitigating it.

Thoughts by