In recent years, the rapid development of artificial intelligence (AI) technology has significantly driven the growth in energy demand for global data centers. Especially, generative AI models like ChatGPT require immense computing power, leading to a sharp increase in data center electricity consumption and triggering an unprecedented energy crisis.

01  Energy Challenges for AI Data Centers

l Rising Electricity Consumption

According to reports by the International Energy Agency, the electricity consumption of global data centers is climbing year by year. In 2022, the total electricity consumption of data centers worldwide approached 460 terawatt-hours (1 terawatt-hour equals 10 billion kilowatt-hours), and this figure is projected to exceed 1,000 terawatt-hours by 2026, equivalent to Japan's annual electricity consumption in 2022. The large-scale application of AI technology is the primary reason for this surge in data center electricity consumption. For instance, the training of the large language model GPT-3 requires a staggering 1,287 megawatt-hours (1 megawatt-hour equals 1,000 kilowatt-hours) of electricity, equivalent to the total energy consumed by 3,000 electric vehicles circling the equator eight times. ChatGPT responds to approximately 200 million requests per day, consuming over 500,000 kWh of electricity, which is 17,000 times the average daily electricity consumption of an ordinary household.

l Threats to Energy Security and Decarbonization Goals

The energy demand of AI data centers not only exacerbates the pressure on electricity supply but also poses a serious threat to decarbonization goals, putting immense strain on the power grid and potentially causing volatility in the energy market. It is predicted that by 2030, the combined electricity demand of electric vehicles and AI will increase the U.S. grid's load by 290 terawatt-hours, making their total electricity consumption roughly equivalent to the national electricity consumption of Turkey, the 18th largest economy in the world.

l Water Consumption and Environmental Issues

Apart from electricity consumption, AI data centers also present significant water consumption challenges. For example, every time ChatGPT engages in a conversation with users involving 25 to 50 questions, it consumes 500 milliliters of water for cooling. This phenomenon of high energy and water consumption is becoming increasingly common with the widespread application of generative AI. Furthermore, the energy consumption of AI contributes to environmental issues, such as greenhouse gas emissions exacerbating climate change and putting pressure on natural resources.

02  The Need for Efficient PSUs and Solutions

The energy consumption of data centers mainly stems from core computing components (including CPUs, accelerators, storage, etc.), precision air conditioners and cooling fans that provide a constant temperature and humidity environment for the machine room, and power supply and distribution systems (including PSUs, distribution cabinets, uninterruptible power supplies (UPS), etc.). As a critical component of data center energy supply, the efficiency of PSUs directly affects data center energy consumption and operating costs.

1. Technologies to Improve PSU Efficiency

² Advancements in Semiconductor Materials: Due to the limited physical properties of traditional silicon-based semiconductors, they do not perform well in the fields of high temperature, high pressure, high frequency and high power. Therefore, gallium arsenide, silicon carbide (SiC), gallium nitride (GaN) and other compound semiconductors came into being. Taking silicon carbide as an example, its high pressure resistance is 10 times that of silicon, high temperature resistance is twice that of silicon, and high frequency capacity is twice that of silicon. In recent years, power semiconductor manufacturers have introduced data center PSU solutions based on silicon carbide or gallium nitride devices. For example, Infineon uses a hybrid switching scheme in a high-power PSU, using silicon, SiC, GaN and other power switching tubes at the same time, and its 3kW PSU scheme can achieve a peak efficiency of 97.5%.

² Optimization of Power Management Technology: By improving power management technology, such as adopting advanced cooling systems and smart grids, the energy utilization efficiency of data centers can be enhanced. For example, PSUs with a totem-pole bridgeless PFC + LLC structure feature an AC-DC bridgeless totem-pole PFC at the frontend and a DC-DC isolated full-bridge LLC converter at the backend, enabling higher energy efficiency and power density.

2. Market Application of Efficient PSUs

As the demand for AI data centers grows, third-generation semiconductor imports are accelerating. Infineon, Onsemi, Nano Semiconductor and other power semiconductor manufacturers have introduced data center PSU solutions based on SiC and GaN devices. These solutions not only improve the energy efficiency and power density of Psus, but also reduce power consumption and CO2 emissions, thereby reducing overall operating costs. Infineon's PSU product roadmap is designed to meet the current and future energy needs of AI data centers. With the introduction of the 8 kW and 12 kW PSU, Infineon has further improved the energy efficiency of its AI data centers. The 12 kW PSU offers advantages for powering future data centers with increased energy efficiency, power density and reliability.

Although quantum computing has significant advantages in reducing AI energy consumption, its practical application faces many challenges, such as the difficulty of maintaining quantum states and the technical hurdle of scaling quantum bit numbers. However, with technological advancements, quantum computing is expected to become one of the important solutions to AI energy consumption issues in the future.

03  Strategies for Addressing the Energy Crisis

Expanding Energy Supply

In response to the rapidly growing energy demand driven by AI, expanding energy supply is a necessary measure. Technology companies are seeking to develop new energy sources, with nuclear power as a stable and carbon-free power source being a key focus. For example, Microsoft has entered into an agreement with Constellation Energy to purchase electricity from the Three Mile Island nuclear power plant for the next 20 years to power its data centers.

Reducing AI Energy Consumption

Apart from expanding energy supply, reducing AI energy consumption is also a crucial approach to addressing the energy crisis. By optimizing AI algorithms and using more efficient machine learning frameworks, redundant computations and model parameters can be reduced, improving computational efficiency and lowering energy consumption. Additionally, adopting more efficient hardware devices, such as dedicated AI chips and servers with higher energy efficiency, can also effectively reduce energy consumption.

Developing Renewable Energy

Developing renewable energy is an important strategy for improving energy utilization efficiency. This includes potential pathways for expanding clean energy sources such as solar photovoltaics and wind energy. By scaling up new carbon-free energy sources, such as geothermal energy, advanced small modular reactors (SMRs), and fusion technology, the carbon emissions of AI data centers can be further reduced.

04  Conclusion

The energy crisis faced by AI data centers is a pressing issue that needs to be addressed. By adopting more efficient PSUs, optimizing power management technology, expanding energy supply, reducing AI energy consumption, and developing renewable energy, this crisis can be effectively tackled. In the future, with continuous technological advancements and the development and utilization of new energy sources, AI data centers will achieve more efficient and sustainable development.

Transceiver ic, Find in Conevo

● The AD8317ACPZ is a 1 MHz to 10 GHz, 55 dB log detector/controller designed for applications such as RF transmitter PA setpoint control and level monitoring, as well as RSSI measurements in base stations, WLAN, WiMAX, and radar systems.

● The RTL8208L-CG is a high-performance Ethernet transceiver chip from Realtek, featuring low power consumption, port-pair loop mode, and advanced functionality for 10Base-T/100Base-TX/FX Ethernet applications, making it a versatile solution for a wide range of communication devices.

● The SI4460-C2A-GMR is a high-performance, low-current ISM EZRadioPRO transceiver designed for sub-GHz frequency bands, offering outstanding sensitivity and low active and standby current consumption, suitable for a wide range of wireless applications.

Website: www.conevoelec.com

Email: info@conevoelec.com