How to Boost Datacenter Efficiency and Power Density Using MLCCs

Technical Overview: Artificial intelligence (AI) workloads are driving rapid increases in datacenter rack power, requiring power architectures that deliver higher efficiency, greater power density, and improved thermal performance. This article explains how multilayer ceramic capacitors (MLCCs) help address these challenges by supporting modern power supply units (PSUs), intermediate bus converters (IBCs), and DC/DC converter topologies used in AI servers.

This article examines the growing role of MLCCs in LLC resonant converters, multilevel power factor correction (PFC) circuits, switched-capacitor converters (SCCs), and hybrid switched-capacitor converters (HSCs). It discusses how characteristics such as dielectric type, capacitance density, equivalent series resistance (ESR), equivalent series inductance (ESL), ripple-current capability, temperature stability, and soft termination influence efficiency, reliability, and power density.

Using examples from TDK's C series MLCC portfolio, this article explores the application of C0G (NP0), X6S, X7T, and X7R dielectric technologies in demanding power conversion applications. Readers will learn how evolving AI server power architectures are influencing capacitor selection and how modern MLCC technologies support more efficient, compact, and reliable datacenter power systems. (Overview courtesy of ChatGPT)

Improved power supply units (PSUs) and intermediate bus converters (IBCs) are key to meeting these demands. These subsystems must support higher power levels while reducing losses and shrinking board footprints. Improvements in capacitor performance and form factor have become key enablers of these goals.

This article outlines the role of capacitors in rack-level power systems and examines the architectural evolution of PSUs and IBCs. It then introduces TDK’s multilayer ceramic capacitors (MLCCs) and shows how they can be applied to address the challenges of AI datacenter power design.

Power trends in datacenters

Datacenter processing loads are straining power delivery architectures from the grid to the processing unit (xPU) (Figure 1). At each stage, high efficiency, low ripple, and robust thermal tolerance are must-haves.

Figure 1: Rising power demand is straining datacenter power architecture, from the grid to the xPU. (Image source: TDK)

The rack and server-level demands are particularly challenging due to rapid increases in compute density. For example, PSU designs are shifting from the conventional 1 kW class to 6 kW or 12 kW and above.

To address this escalating demand, designers are seeking ways to reduce losses and manage heat within the PSU while minimizing the printed circuit board (pc board) footprint. Improvements in the performance and form factor of passive components, such as capacitors, are key enablers for meeting these goals.

TDK’s broad lineup of MLCCs is a prime example of capacitors well-suited to meet datacenter power delivery requirements. These capacitors excel in capacitance density while delivering high performance and robust thermal tolerance across various power stages as architectures evolve.

Changes in DC/DC circuit topology for increased PSU power

The DC/DC conversion stage performs the bulk power conversion at high voltage and high switching frequency, so minimizing losses here is critical to overall efficiency. Inductor-inductor-capacitor (LLC) resonant converters are the dominant topology for this stage. As rack power increases, designers are migrating to phase-parallel (interleaved) designs and series/parallel power-block configurations. These designs use large inductor-capacitor (LC) resonant tanks to distribute current and heat while increasing power capacity (Figure 2).

Figure 2: The DC/DC converters in datacenter PSUs require high-voltage resonant capacitors. (Image source: TDK)

Polypropylene film capacitors were the traditional choice for this application. As switching frequencies rise, however, resonant tanks increasingly adopt MLCCs for their stability, smaller footprint, and lower losses.

TDK’s high-voltage Class 1 C0G MLCCs are a good example. The C0G temperature coefficient means that capacitance varies by 0 ±30 parts per million per °C (ppm/°C). This rating is also known as negative-positive-zero (NP0) because capacitance remains virtually unchanged across the operating temperature range. This helps resonant tanks remain in their most efficient operating region even as the operating temperature changes during heavy-load transients, which are common in modern AI datacenters.

Also notable is the family’s minimum quality factor (Q) of 1000, indicating extremely low equivalent series resistance (ESR) and correspondingly low dissipating loss. In practical terms, this corresponds to a dissipation factor of 0.1% or less, resulting in lower conversion losses than a comparable film capacitor. In addition, the TDK Class 1 C0G MLCCs are packaged in a compact 3.2 × 2.5 millimeter (mm) (EIA 1210) case that occupies significantly less pc board space than a comparable film capacitor.

Capacitors in this MLCC family span 630 V to 1250 V and 2.2 to 10 nanofarads (nF), giving designers options for different LLC architectures. All family members are rated from −55°C to +125°C, providing the thermal tolerance needed for high-power PSU enclosures.

A typical example is the C3225C0G3B103J250AC, rated at 1250 V and 10 nF. Its temperature variation curve is representative of the family, with less than 1% deviation across the supported temperature range (Figure 3).

Figure 3: The C3225C0G3B103J250AC exhibits less than 1% capacitance variation (no bias, Y axis set to 0.0000000095 to 0.0000000105). (Image source: TDK, modified by Kenton Williston)

Changes in PFC topology for increased PSU power

At the higher power levels now typical of datacenter PSUs, the PFC front end is migrating to multilevel topologies that use flying capacitors (Figure 4). These topologies create intermediate voltage levels that reduce stress on the switching semiconductors, reducing switching losses and enabling the use of lower-rated components.

Figure 4: Multilevel PFC topologies use flying capacitors to reduce the voltage stress on the switching semiconductors. (Image source: TDK)

Metalized polypropylene film capacitors have traditionally served this role. As power density increases, MLCCs offer key advantages in capacitance density, equivalent series inductance (ESL), and temperature capability. In a three-level configuration, each flying capacitor sees roughly half the DC bus voltage, making 450 V MLCCs a good match.

TDK’s C series mid/high-voltage Class 2 MLCCs (X6S/X7T) are well suited to this role, offering 450 V parts in a 5.7 mm × 5.0 mm (EIA 2220) case with capacitance from 0.68 µF to 2.2 µF. Within this family, the X6S and X7T dielectrics trade off temperature rating for capacitance. The X6S parts hold 2.2 µF in the same footprint and are rated to +105°C, prioritizing capacitance density over extreme-temperature tolerance. The X7T parts are rated to +125°C but top out at 1 µF, making them suitable for the hot zones near switching devices.

The C5750X7T2W105K250KE is a typical high-temperature part, offering 1 µF capacitance at 450 V and +125°C. A key feature of this part is its soft termination (Figure 5). Soft termination adds a flexible conductive resin layer that absorbs pc board flex and thermal stress, reducing strain on the ceramic layers. This robustness is particularly valuable for larger parts, such as 2220-size Class 2 MLCCs.

Figure 5: Soft termination adds a flexible resin layer to relieve mechanical stress and thermal shock. (Image source: TDK)

At the other end of the spectrum is the C5750X6S2W225K250KA. This Class 2 450 V MLCC has a much higher capacitance of 2.2 µF, with tradeoffs including a +105°C rating and standard termination. This part is suitable for PFCs with less severe thermal stress.

Converters for high efficiency and high density IBCs

The IBC steps down the 48 V distribution bus to the 12 V rail that feeds the server board. This stage must deliver high efficiency and high power density, given the extreme space constraints at the board level.

To meet these goals, designers are migrating from the conventional LLC to the switched-capacitor converter (SCC) or, more recently, to a hybrid switched-capacitor converter (HSC). Both approaches replace bulky transformers with more efficient capacitors as the primary energy-handling components. This increases the capacitor count in the IBC (Figure 6), making capacitance density a central design constraint. This shifts the capacitor from a supporting component to a primary energy-transfer element.

Figure 6: The SCC and HSC use more capacitors than a traditional LLC. (Image source: TDK)

MLCCs address the density requirement in two ways. First, their high capacitance density allows designers to place many devices in parallel, increasing total capacitance while distributing current and thermal load across the board. Second, their low ESR and ESL make them robust against high-frequency ripple and switching transients, thereby helping limit heat generation.

Flying capacitors play a particularly important role because they are used in both SCCs and HSCs. TDK’s general-voltage X6S/X7R/X7S 50 V family serves these nodes well. It offers capacitances from 220 picofarads (pF) to 22 µF in cases measuring 0.5 mm × 1.0 mm (EIA 0204) to 7.5 mm × 6.3 mm (EIA 3025). With these tiny packages, many capacitors can be paralleled within a given pc board area.

The C2012X7R1H475K125AE is a compact option, offering 4.7 µF in a 2.0 mm × 1.25 mm (EIA 0805) case. Despite its small size, it can handle roughly 3 A of ripple current with only a 20°C rise at switching frequencies from 500 kHz to 1 MHz. Thus, a handful of these devices in parallel can comfortably handle the aggregate ripple (Figure 7).

Figure 7: The ripple-current capability of the C2012X7R1H475K125AE is approximately 3 A for a 20°C temperature rise. (Image source: TDK)

The SCC/HSC input capacitor also merits consideration. It sits on a 48 V rail, so it needs a 100 V rating to cover the rail voltage plus switching margin. TDK’s mid-voltage X6S/X7R 100 V family fills this role, with package sizes from 1.6 mm × 0.8 mm (EIA 0603) to 5.7 mm × 5.0 mm (EIA 2220) and capacitance up to 10 µF.

The C3216X6S2A106K160AC is an excellent example, packing 10 µF into a compact 3.2 mm × 1.6 mm (EIA 1206) case. This part is notable for its charge density (µF × V per unit volume), which is significantly higher than many electrolytic capacitors used in similar applications and exceeds that of any other part discussed in this article.

It also offers a very low ESR of less than 3 mΩ from 100 kHz to 2 MHz (Figure 8), which is orders of magnitude lower than that of a comparable conventional aluminum electrolytic capacitor. This allows the MLCC to filter the switching-frequency input ripple with minimal voltage drop and self-heating.

Figure 8: The C3216X6S2A106K160AC keeps its ESR under 3 mΩ across the 100 kHz to 2 MHz range. (Image source: TDK)

Conclusion

As AI rack power continues to climb, component selection becomes increasingly important. High-voltage and high-density MLCCs enable designers to improve efficiency, reduce footprint, and support the evolving architectures used in modern PSUs and IBCs. The broad lineup of TDK offerings gives designers the flexibility needed to address these requirements across a variety of power system architectures.