Artificial intelligence (AI) is currently the hottest direction in technology development. However, AI applications rely on vast data centers and computational power, which also means consuming a significant amount of electricity. Improving power conversion efficiency and avoiding energy waste are crucial issues for sustainable human development. This article will introduce the power challenges faced by AI data centers and the MOSFET solutions provided by onsemi.
Power challenges faced by AI data centers
Electricity is at the core of modern society and economic operations, and with the increasing demand for electric vehicles and artificial intelligence applications, its importance will only continue to grow. Power generation is currently the largest source of carbon dioxide (CO2) emissions worldwide, but it can also lead the transition to net-zero emissions through the rapid expansion of renewable energy sources such as solar and wind power. Ensuring that consumers can access electricity safely and affordably while reducing global carbon dioxide emissions is one of the core challenges of the energy transition.
According to the International Energy Agency (IEA), data centers consumed about 2% of all electricity in 2022, which is approximately 460 terawatt-hours (TWh). With the rise of energy-hungry applications such as cryptocurrency and artificial intelligence/machine learning (AI/ML), this number is expected to rise rapidly. This surge in power consumption is based on the deployment of high-performance graphics processing units (GPUs) in these technologies. The IEA predicts that data centers will consume at least 650 TWh by 2026, but consumption exceeding 1,000 TWh is not out of the question.
The growth rate in the field of artificial intelligence is quite astounding. ChatGPT reached 1 million users within the first five days and 100 million users within the first two months, significantly outpacing the growth rates of TikTok and Instagram. Training GPT-4, with its 1.7 trillion parameters and using 13 trillion tokens, requires 25,000 NVIDIA A100 GPUs, with each server consuming about 6.5 kW. According to OpenAI, this training took 100 days, consumed 50 GWh of energy, and cost $100 million.
Early data centers converted grid voltage to 12V centrally and bussed it to servers where logical level conversions (3.3/5V) were completed. As power requirements increased, this method resulted in too much loss. The bus voltage was increased to 48V, reducing the current by 4x and losses by 16x.
As processor voltages dropped below 3.3V to sub-volt levels, multiple voltage rails were needed at relatively high power. This led to a two-stage conversion process where a DC-DC converter (known as an intermediate bus converter (IBC)) converts 48V to a 12V local bus before performing lower voltage conversions.
Efficient power conversion needed for AI data centers
The power conversion requirements of AI data centers are particularly important due to their high-performance computing and extensive data processing demands. AI data centers need to handle large amounts of data and complex computational tasks, which means they require high-efficiency and high-density power conversion systems. Efficient power converters can reduce energy losses, thereby improving the overall system performance and efficiency.
The operation of data centers requires highly stable and reliable power supplies. Power converters must provide stable voltage and current under various load conditions to ensure the normal operation of servers and other equipment. In addition, efficient power conversion systems can reduce heat generation, though effective thermal management is still necessary. Optimized thermal designs help keep the system's temperature within a safe range, thereby extending the lifespan of equipment and improving performance.
As AI applications rapidly develop, the demands on data centers continue to grow. Power conversion systems need to have good scalability to adapt flexibly to future expansion needs. Especially given the substantial energy consumption of AI data centers, energy efficiency management is key to reducing operational costs and environmental impact. Efficient power converters can significantly reduce energy consumption and improve energy utilization efficiency.
To ensure the high availability of data centers, power conversion systems typically need to be designed with redundancy to handle potential power failures. Redundant designs can provide backup power, switching quickly when the primary power source fails, ensuring continuous system operation. Moreover, with the growing environmental awareness, more data centers are starting to incorporate green energy sources, such as solar and wind power. Efficient power conversion systems can better integrate these renewable energies, improving overall energy efficiency and reducing the carbon footprint.
During the power conversion process, power losses are an inevitable phenomenon. These losses constitute wasted energy, incur costs, and generate heat that requires space and further expense to manage. When operating hyperscale AI data centers, which may require 120 kW of rack power, the efficiency of converting grid power to GPU voltage is about 88%, resulting in 15 kW of waste heat that must be managed through liquid cooling.
Efficiency and power density (which go hand in hand) are the key terms in server power design. The energy from the main grid must be converted to useful power with minimal losses. To achieve this, topologies are continuously evolving, with technologies such as synchronous rectification being developed, and MOSFETs replacing lossy diodes in rectifiers.
Enhancing topology is only half the battle for success; to optimize efficiency, all components must be as efficient as possible, especially the MOSFETs that are critical to the conversion process. MOSFETs are not lossless devices; they incur losses during conduction and switching. As server power supplies shift to higher frequency operations to reduce size, switching losses become a key focus for improvement.
Efficient onsemi PowerTrench® MOSFETs
Silicon MOSFETs control the current between the source and drain terminals through gate voltage. Due to their efficiency, speed, and power handling capabilities, they are widely used in power amplifiers, voltage regulators, and switching circuits. onsemi's low to medium voltage T10 PowerTrench® MOSFETs reduce switching and conduction losses through the latest shielded gate trench technology, resulting in significantly lower Qg and sub-1mΩ RDS(ON). The industry's leading soft recovery body diode reduces ringing, overshoot, and noise as well as Qrr losses, balancing performance and recovery in fast switching applications. Compared to earlier devices, these new MOSFETs can reduce switching losses by up to 50% and conduction losses by over 30%.
onsemi's new 40V and 80V T10 PowerTrench devices offer best in class RDS(ON). The NTMFWS1D5N08X (80V, 1.43mΩ, 5mm x 6mm SO8-FL package) and the NTTFSSCH1D3N04XL (40V, 1.3mΩ, 3.3mm x 3.3mm source-down dual-cool package) feature best-in-class figure of merit (FOM), making them suitable for power supply units (PSU) and intermediate bus converters (IBC) in AI data center applications. T10 PowerTrench MOSFETs meet the stringent Open Rack V3 efficiency specification, which requires efficiencies of 97.5% or higher.
Low/medium voltage MOSFETs with better performance
The low/medium voltage MOSFETs introduced by onsemi, specifically the NTMFWS1D5N08X, are power, single, N-channel MOSFETs with an STD gate, using the SO8FL-HEFET package. They support 80V, 1.43mΩ, and 253A. This T10 80V MOSFET is among the best-in-class products in the 80V market, making it an optimal solution for cloud power, 5G telecom, other PSU applications, DC/DC, and industrial applications. It offers better performance and improves system efficiency and high-power density, though it has lower performance features.
The NTMFWS1D5N08X features improvements in FOM, Rsp, and power density, enhancing performance and reducing costs. Its lower Rsp, low Qg/Qgd, and lower Qgd/Qgs can increase overall efficiency by minimizing driver losses. The low RDS(ON) minimizes conduction losses, while the lower Qoss and Qrr improve switching losses. A softer recovery diode and lower Qrr reduce ringing, overshoot, and noise, providing robustness and excellent Unclamped Inductive Switch (UIS) to improve avalanche ruggedness in fast-switching applications. These devices are Pb-free, halogen/BFR-free, and RoHS compliant.
The NTMFWS1D5N08X is suitable for DC-DC and AC-DC synchronous rectification (SR) and primary switches in isolated DC-DC converters and motor drives. Common end products include telecom power, cloud power, server power, data centers, motor drives, solar power, and uninterruptible power supplies (UPS).
Another low/medium voltage MOSFET from onsemi is the NTTFS2D1N04HL, an N-channel shielded gate PowerTrench® MOSFET supporting 40V, 150A, and 2.1mΩ. This N-channel medium voltage MOSFET is produced using advanced PowerTrench® technology incorporates with shielded gate technology. This process is optimized to minimize on-state resistance while maintaining excellent switching performance.
The NTTFS2D1N04HL, utilizing shielded gate MOSFET technology, has a maximum RDS(on) of 2.1mΩ at VGS = 10V and ID = 23A, and 3.3mΩ at VGS = 4.5V and ID = 18A, lowers switching noise and EMI. It features a robust package design that meets MSL1, is 100% UIL tested, and RoHS compliant. The NTTFS2D1N04HL is versatile, suitable for many different applications, with common end products including DC-DC power supplies and medium voltage synchronous buck converters.
Conclusion
In today's rapidly evolving era of artificial intelligence, improving power conversion efficiency is crucial. High-efficiency power conversion technology can not only meet the demands of AI applications for high-performance computing and extensive data processing but also significantly reduce energy consumption and operating costs, thereby achieving more sustainable and environmentally friendly goals. By continuously innovating and adopting advanced power management solutions, we can strike a balance between performance and efficiency, ensuring the sustained development and maximal benefit of AI technology in various fields. Therefore, investing in improving power conversion efficiency is not only a technological advancement requirement but also a vital part of driving the AI revolution.
The low to medium voltage PowerTrench® MOSFET product series launched by onsemi, with its excellent performance, can be applied to data centers for AI applications, providing outstanding power conversion efficiency and making it one of the ideal choices for related applications.
