How to Provide Effective Power Management for AI Datacenters

The rise of artificial intelligence (AI) and machine learning (ML) has created unprecedented power demands. The next generation of datacenters faces significant challenges in power management, efficiency, and reliability. Traditional power solutions often struggle to meet these demands at the level of individual components and overall datacenter infrastructure management (DCIM). Advanced power components and integrated monitoring solutions offer a comprehensive approach to meeting these challenges.

For example, hybrid capacitor technology provides stable power delivery; ultra-low equivalent series resistance (ESR) solutions afford efficiency in high-current power conversion; high-accuracy resistors enable precision power monitoring; and wireless integration provides comprehensive power management.

This article explores how these elements help create robust power management systems for AI-driven datacenters. It then introduces Panasonic solutions across all four areas and demonstrates their application in modern datacenter environments.

Efficient datacenter power delivery with hybrid capacitor technology

Modern datacenters require extensive power conversion. It is common for them to need hundreds of kilovolts AC (kVAC) from the grid. This voltage is first stepped down to tens of kVAC for distribution across the datacenter campus. It is then further converted to hundreds of VAC for distribution to the equipment racks.

At the rack level, AC power is converted to direct current, usually 12 volts DC (VDC), to meet the requirements of the IT equipment. Finally, within each piece of equipment, the voltage is further regulated to lower levels, often between 1.1 and 5 volts, to power the individual components such as processors and memory modules.

Each step in this chain introduces losses that can significantly impact the datacenter's overall efficiency. Data center power designers are increasingly adopting wide-bandgap (WBG) semiconductors like gallium nitride (GaN) to minimize losses in the later conversion stages. Compared to traditional silicon (Si), WBG devices achieve superior efficiency through higher switching frequencies and lower conduction losses.

However, the capacitor technology used in these converters presents significant design challenges. Power system designers have traditionally had two proven capacitor technologies: conventional aluminum electrolytic capacitors, which feature low leakage current, and polymer capacitors, which have outstanding ESR characteristics. Panasonic’s EEH Series hybrid aluminum electrolytic capacitors (Figure 1) present a third option that combines the strengths of both to minimize losses due to leakage current and ESR.

Figure 1: The EEH series hybrid aluminum electrolytic capacitors minimize losses due to leakage current and ESR. (Image source: Panasonic)

Hybrid capacitors have other advantages, including enhanced reliability through open-circuit failure modes and maintaining their rated capacitance at much higher frequencies than traditional designs. While conventional capacitors begin to lose effectiveness at frequencies in the tens of kilohertz (kHz), hybrid capacitors retain their performance at frequencies approaching 1 megahertz (MHz). This higher operating frequency enables the use of smaller capacitors, allowing designers to create more compact converters or free up board space for additional features.

A typical hybrid capacitor example is the EEH-ZA1V151P. This 150 microfarad (µF), 35 volt device maintains a low ESR of 27 milliohms (mΩ), has an operating temperature range of -55°C to approximately +105°C, and features a lifespan of 10,000 hours (hrs) (at +105°C). Its suitability for datacenter applications is demonstrated in the EVLMG1-250WLLC DC/DC converter evaluation board from STMicroelectronics (Figure 2). This GaN board achieves power densities of 20 watts per cubic inch (W/in.³) at better than 92% efficiency.

Figure 2: The EVLMG1-250WLLC GaN DC/DC converter evaluation board demonstrates the potential of the hybrid capacitor. (Image source: STMicroelectronics)

Advantages of low-ESR capacitors for high-density, high-efficiency power delivery

The trend toward high-power-density DC/DC converters in datacenters creates unique thermal management challenges. The increasing power density and reduced component area can dramatically raise operating temperatures.

Minimizing a capacitor’s ESR can partly address these thermal challenges. Since power loss follows the I²R relationship, reducing the resistance directly decreases power loss and, consequently, heat generation. This makes low ESR crucial for maintaining safe operating temperatures in compact designs.

However, even the most efficient capacitors can experience high operating temperatures due to their environment. Thus, selecting a capacitor that can withstand the heat of a tightly packed datacenter is essential. Figure 3 shows a selection chart that factors in the operating temperature, among other considerations.

Figure 3: Shown is a selection guide for hybrid capacitors based on ripple current, capacitance, size, and operating temperature. (Image source: Panasonic)

While the high switching frequencies enabled by GaN technology allow for smaller packages, the capacitor technology must maintain adequate capacitance to handle high ripple currents. With capacitance options from 47 μF to 680 μF and the ability to handle up to 2.3 amperes (A) at 100 kHz, the EEH-ZL Series hybrid capacitors address these challenges. They also have guaranteed operation to +135°C and an ESR down to 14 mΩ.

An example is the EEH-ZL1E681P 680 μF capacitor, which has an ESR of 14 mΩ and a package diameter of 10.0 mm.

Using high-precision resistors for precise power monitoring

DC/DC converters in datacenter applications require highly accurate feedback for power control. This is especially critical in GaN-based designs, where even minor errors in duty-cycle feedback can result in dangerous overvoltage or overcurrent conditions.

While various current-sensing technologies exist, shunt resistors are particularly appealing for the space-constrained environments of servers, storage infrastructure, and power supplies. However, modern designs' high power density creates significant challenges for resistive current sensing.

The primary challenge lies in thermal stability. Resistance values can drift significantly as operating temperatures change, potentially compromising measurement accuracy. This makes the thermal coefficient of resistance (TCR) a critical specification. It must be as low as possible to maintain measurement precision across the wide temperature ranges encountered in datacenter operations.

The Panasonic ERA-8P series resistors (Figure 4) address these challenges through several innovative features:

• An ultra-low TCR of ±15 × 10^-6 per degree Kelvin (K) achieved through precision thin-film processing
• A stress-reducing soft resin layer beneath the resistor that minimizes solder crack formation during thermal cycling
• A smooth alumina substrate surface that ensures uniform resistive film thickness
• A long, fine serpentine resistance pattern that disperses current load concentration, providing industry-leading electrostatic discharge (ESD) resistance

Figure 4: The ERA-8P series resistors are designed for high thermal stability. (Image source: Panasonic)

The ERA-8PEB1004V demonstrates these capabilities with specifications suited to datacenter power monitoring:

• A high limiting element voltage of 500 V at 1 MΩ for monitoring high-voltage power rails
• A 0.25 W power rating that ensures minimal power loss
• A wide operating temperature range of -55°C to +155°C
• A superior electrostatic discharge (ESD) resistance for reliable operation in high-power environments

Using Wi-Fi to monitor power efficiency

DCIM faces growing complexity as AI workloads drive the deployment of more servers, storage systems, and power supply units. While monitoring power consumption across these systems is crucial for optimizing efficiency, traditional wired monitoring solutions add cost, complexity, and cable-management challenges that only compound as facilities scale.

Wireless monitoring offers an elegant solution to these challenges. It enables real-time power management through voltage, current, and temperature measurements without the overhead of additional cabling. This approach provides greater flexibility for scaling operations up or down without reconfiguring the physical connections.

However, wireless modules for datacenter applications must address several stringent requirements:

• Maintain reliable connectivity in environments with numerous obstacles and potential interference sources
• Minimize power consumption to maintain overall efficiency gains
• Fit within compact form factors to integrate with existing equipment
• Provide robust security features to protect sensitive datacenter information

The Panasonic ENW-49A01A3EF PAN9320 Wi-Fi module (Figure 5) addresses these challenges through its comprehensive feature set:

• 2.4 GHz operation provides superior penetration through datacenter obstacles while ensuring broad compatibility through support for 802.11b/g/n standards.
• Power efficiency is maintained through a minimum transmit (Tx) power consumption of 430 milliamperes (mA) for transmit (Tx) and 160 mA for receive (Rx) in 802.11b mode.
• A compact 29.0 mm × 13.5 mm × 2.66 mm surface-mount design simplifies integration.
• Built-in security features, such as TLS/SSL, HTTPS, and WPA2, protect sensitive information.

These capabilities enable datacenter operators to implement comprehensive power monitoring while minimizing the physical and operational overhead typically associated with such systems.

Figure 5: The ENW-49A01A3EF provides a comprehensive 2.4 GHz Wi-Fi solution for effective DCIM. (Image source: Panasonic)

Conclusion

The demands of AI workloads require a rethink of their power infrastructure, from individual component selection to facility-wide monitoring systems. Panasonic’s portfolio of hybrid capacitors, ultra-low ESR technology, precision resistors, and wireless connectivity provides datacenter operators with the tools they need to build and maintain efficient, scalable power systems to support next-generation AI applications.