Driven by emerging applications such as cryptocurrency and artificial intelligence/machine learning (AI/ML), the energy consumption of data centers is enormous and will rapidly grow to meet user demands. According to the latest report from the International Energy Agency (IEA), the power consumption of data centers will reach 460 TWh (terawatt hours) in 2022, accounting for approximately 2% of the global total electricity consumption. In the United States, one-third of the world's data centers consume 260 TWh of electricity, accounting for 6% of the total electricity consumption.

Predicting the future is challenging, depending on how many highly power consuming graphics processing units (GPUs) are deployed to meet the demands of artificial intelligence technology, as well as further increasing air conditioning to lower the temperature in data centers. The International Energy Agency's report shows that by 2026, the power consumption of data centers will increase to at least 650 TWh (40%), but may also reach as high as 1050 TWh (128%).

Data centers support the trend of artificial intelligence

Artificial intelligence is an extremely power consuming technology, and the data centers that support its operation need to have sufficient computing power and power transmission capabilities.

A recent study by the Swedish RISE research institute clearly demonstrates the significant changes brought about by the rapid adoption of this technology. For example, ChatGPT reached 1 million users within just five days of its launch in November 2022. They had 100 million users within two months, while TikTok reached the same user base in nine months and Instagram took two and a half years.

As a reference, conducting a search on Google only requires 0.28 Wh, which is equivalent to keeping a 60W light bulb on for 17 seconds.

In contrast, training GPT-4 requires 1.7 trillion parameters and 13 trillion tokens (word fragments), which is a completely different proposition. To achieve this, multiple servers with 25000 NVIDIA A100 GPUs are required, each with a power consumption of approximately 6.5 kW. OpenAI stated that the training took 100 days, consumed approximately 50 GWh, and cost $100 million.

Obviously, artificial intelligence will greatly change the game rules of data centers, requiring computing power and energy consumption levels far beyond any level we have seen so far.

48V architecture for data centers

Early data centers adopted a centralized power architecture (CPA), which centrally converted the main power (grid) voltage to 12V (bus voltage), distributed it to each server, and locally converted it to 5V or 3.3V logic levels using relatively simple converters.

However, with the increase in power demand, the current (and related losses) on the 12V bus became unacceptably high, forcing system engineers to switch to a 48V bus arrangement. According to Ohm's law, if the current decreases by four times, the loss decreases by four times the square. This configuration is called Distributed Power Architecture (DPA).

At the same time, the voltage of the processor and other components is constantly decreasing, eventually dropping to the sub volt level, requiring multiple secondary voltage rails. To solve this problem, second-order conversion technology is adopted, which converts 48V voltage to 12V bus through DC-DC converter (known as intermediate bus converter IBC), and then outputs other voltages from the 12V bus as needed.

The demand for high-efficiency MOSFETs

The power loss within the data center poses a challenge for operators. Firstly, and most notably, they are paying for the electricity that does not contribute to the operation of the servers. Secondly, any wasted energy will be converted into heat, which must be managed. Due to the power demand of ultra large scale AI servers reaching up to 120 kW (which is bound to increase over time), even at 50% load, with a peak efficiency of 97.5% and a loss of 2.5%, each server would waste 1.5 kW of electricity, equivalent to a fully operational electric heater.

Handling heat may require heat dissipation measures such as radiators or fans in the power conversion system. These measures will increase the size of the power supply, occupy space that could have been used for more computing power, and in terms of fans, consume electricity and increase costs. Due to the strict temperature control required within the data center, excessive losses can also raise the ambient temperature, which means more air conditioning is needed to cool down. This is both capital expenditure and operating cost, while also occupying space.

Obviously, efficiently converting the main (grid) voltage into the voltage required to power artificial intelligence GPUs and other devices is of great benefit to data center operators.

Therefore, over the years, people have done a lot of work in power topology, introducing technologies such as totem pole PFC (TPPFC) in the front-end PFC stage to improve its efficiency. In addition, in order to improve efficiency, diode rectifiers have been replaced by MOSFETs and technologies such as synchronous rectification have been introduced.

Optimizing the topology structure is only half of it. To optimize efficiency, all components must also be as efficient as possible, especially MOSFETs that are crucial for the conversion process.

When MOSFETs are used for switching power conversion, there are mainly two forms of losses: conduction loss and switching loss. The conduction loss is caused by the resistance (RDS (ON)) between the drain and source, which exists continuously when current flows. Switching losses are caused by a combination of gate charge (Qg), output charge (QOSS), and reverse recovery charge (Qrr), which are replenished during each switching cycle. Due to the current trend of increasing the switching frequency to reduce the size of magnetic components, this loss will become quite significant as the supplementary frequency increases.

Obviously, the lower the conduction loss and switching loss of a specific MOSFET, the higher the overall conversion efficiency of the power system.

Introduction to PowerTrench T10 MOSFET

Synchronous rectification has now become a key technology in all high-performance, high current, and low-voltage power conversion applications, especially in the application of data center servers. In this application, several MOSFET parameters including RDS (ON), Qg, QOSS, and Qrr directly affect conversion efficiency, and device manufacturers are striving to find ways to reduce these effects.

Anson's PowerTrench T10 MOSFET adopts a new shielded gate channel design, achieving ultra-low Qg values and RDS (ON) below 1mOhm. The latest PowerTrench T10 technology not only reduces ringing, overshoot, and noise, but its industry-leading soft recovery diode also lowers Qrr. This achieves a good balance between on resistance performance and recovery characteristics, while also enabling low loss fast switching with good reverse recovery characteristics.

Overall, the parameter improvements of the PowerTrench T10 device have improved the efficiency of solutions for medium and low voltage, high current switching power supplies. Normally, switch losses can be reduced by up to 50% compared to the previous generation devices, while conduction losses can be reduced by 30% -40%.

Anson Mei has launched the 40V and 80V series products with PowerTrench T10 technology. NTMFWS1D5N08X (80V, 1.43m Ω, 5mm x 6mm SO8-FL package) and NTTFSCH1D3N04XL (40V, 1.3m Ω, 3.3mm x 3.3mm source cooling package) provide excellent efficiency (FOM) for power supply units (PSU) and intermediate bus converters (IBC) in artificial intelligence data center applications. They have achieved a PSU efficiency of 97.5% and an IBC efficiency of 98% as required by the Open Rack V3 specification.

epilogue

The artificial intelligence revolution has arrived, and no one can fully determine what it means for the future power transmission needs of data centers. However, it is certain that a series of new challenges have emerged. The scarcity of real estate resources and the limitations of the power grid make it difficult to find new locations with sufficient capacity. The overall surge in power demand in critical IT areas has placed a heavy burden on electricity costs. To meet these demands, data center owners not only need to build new facilities, but also push existing facilities to the limit, striving to achieve high-density configurations of megawatts per square foot.

As the power level is bound to exceed 100 kW, power conversion will become a key focus to achieve efficient operation, ensure heat dissipation, reliably increase power density, and save space in narrow modern data centers.

Anson's PowerTrench T10 technology provides industry-leading RDS (ON), higher power density, reduced switching losses, and better thermal performance, thereby reducing overall system costs. Innovative power semiconductor technologies such as PowerTrench T10 will become a key component of the future.

This is reported by Top Components, a leading supplier of electronic components in the semiconductor industry

They are committed to providing customers around the world with the most necessary, outdated, licensed, and hard-to-find parts.

Media Relations

Name: John Chen

Email: salesdept@topcomponents.ru