gesponsertPassive Components Why AI Servers Need a New Generation of MLCCs

Gesponsert von

Taiyo Yuden Co. Ltd.

AI accelerators are now exceeding the 1 kW-per-GPU threshold, while server racks are approaching the megawatt level. An examination of the Power Distribution Network (PDN) of modern AI servers reveals why multilayer ceramic capacitors (MLCCs) are becoming increasingly important.

It is no longer sufficient to simply provide enough electrical power. Instead, this energy must be delivered to the processor with the lowest possible losses while simultaneously coping with extremely fast load transients. Every unnecessary watt of power loss not only increases the energy consumption of the data center but also makes cooling significantly more challenging.

At the center of this development is the Power Distribution Network (PDN)—the complete power delivery chain from the power supply through the voltage regulators to the processor itself. Power supplies and DC/DC converters tend to dominate the conversation, yet passive components are gaining ground in importance. Multilayer ceramic capacitors (MLCCs) are a case in point: by stabilizing supply voltage even under highly dynamic load transients, they allow modern AI processors to consistently reach their maximum computing performance.

For manufacturers of passive components, this trend represents a fundamental technological shift. The market demands MLCCs with significantly higher capacitance, lower parasitic inductance, and increasingly compact form factors. At the same time, new integration concepts are becoming more important to utilize the limited space available on the PCB and within semiconductor packages as efficiently as possible. These requirements form the basis of TAIYO YUDEN's current development strategy.

Power Delivery Moves Closer to the Processor

For many years, supply voltages were generated by high-performance voltage regulators located on the motherboard and then routed to the processor through PCB traces. As long as processors consumed only a few hundred watts, this architecture posed few challenges.

The conditions for AI accelerators, however, look very different. Their core supply voltage typically ranges between 0.7 V and 0.9 V, while power levels exceeding 1 kW push currents well past 1,000 A. Under these conditions, even minimal trace resistance produces measurable voltage drops and substantial power losses.

Another phenomenon further complicates power delivery. AI processors can change their power consumption within just a few nanoseconds. As neural networks execute different computational workloads or multiple compute units become active simultaneously, extremely fast load transients occur. The power delivery system must respond virtually instantaneously. Otherwise, the supply voltage temporarily drops below its target level, negatively affecting both system stability and computing performance.

For this reason, the architecture of modern server power delivery is undergoing a fundamental transformation. The final stage of voltage regulation is moving progressively closer to the processor.

TAIYO YUDEN describes this evolution in three stages:

• Motherboard Voltage Regulator (MBVR): Voltage regulation is located entirely on the motherboard.

• Vertical Power Delivery (VPD): The power delivery circuitry moves closer to the processor, shortening current paths and reducing losses.

• Integrated Voltage Regulator (IVR): The final voltage conversion stage is integrated directly into the processor package or its immediate vicinity.

The shorter the current path between the voltage regulator and the processor, the lower the resistive losses become. At the same time, transient response improves dramatically because energy reaches the processor much faster. This evolution is now considered one of the key enabling technologies for future AI server architectures.

However, this progress introduces a new challenge. As power delivery moves closer to the processor, the available installation space becomes increasingly limited. Consequently, smaller components become the only viable solution.

Why MLCCs Are Essential for AI Servers

From the outside, multilayer ceramic capacitors appear to be among the smallest components on a printed circuit board. Their importance for modern AI servers, however, is enormous.

While voltage regulators generate the required supply voltage, MLCCs temporarily store electrical energy directly adjacent to the processor and can release it again within just a few nanoseconds whenever required.

This behavior can be compared to a small local energy reservoir. If the supply voltage temporarily drops because of a sudden increase in current demand, the MLCCs located immediately next to the processor instantly release their stored charge. This keeps the supply voltage within acceptable limits until the voltage regulator has had sufficient time to adjust its output accordingly.

How Modern MLCCs Achieve Higher Capacitance

A capacitor consists of two conductive electrodes separated by an insulating layer, the dielectric. When voltage is applied, opposite charges build up on the electrodes, and the resulting electric field stores energy. The moment the supply voltage briefly dips, this stored energy discharges again within just a few nanoseconds — fast enough to buffer even the rapid load transients of AI accelerators.

This is precisely what makes MLCCs indispensable for modern processors. Voltage regulators handle the average supply voltage, but it's the MLCCs that deliver the energy needed to keep pace with the extremely fast load transients constantly occurring in AI accelerators. In effect, they act as local, high-speed energy reservoirs.

A single capacitor would not be sufficient for this purpose. Instead, an MLCC consists of numerous stacked layers of ceramic dielectric material and nickel electrodes. The ceramic, typically barium titanate (BaTiO₃), forms the dielectric, while the nickel layers serve as electrodes. Each individual layer functions as a small capacitor. By stacking hundreds or even thousands of these layers, manufacturers achieve very high capacitance within an extremely compact component.

The achievable capacitance is primarily determined by four factors: the electrode area, the number of active layers, the dielectric constant of the ceramic material, and the thickness of the dielectric layer.

Since increasing the package size is generally not an option, current development focuses on realizing increasingly finer structures within the same footprint. Thinner ceramic layers combined with a greater number of active layers enable substantially higher capacitance without increasing component dimensions.

High-Density MLCCs Meet Increasing Power Density Requirements

The power consumption of modern AI processors continues to rise, while their supply voltages are simultaneously decreasing to well below one volt. For system designers, this creates a significant challenge: ever-higher currents must be delivered within an increasingly confined space, while sufficient local energy storage must still be maintained.

To address these demands, TAIYO YUDEN has developed a dedicated roadmap for AI server applications. Its portfolio includes High-Density MLCCs from the X6 and X7 series, optimized for the various voltage rails found in modern server architectures. These products range from high-capacitance output capacitors for supply voltages between 0.6 V and 1.2 V to MLCCs designed for intermediate bus voltages and higher-voltage input filtering.

The objective is to provide sufficient capacitance despite the severe space constraints of next-generation AI platforms while ensuring the stability of the entire Power Distribution Network.

Embedded MLCCs Create New Opportunities for System Design

As power delivery circuitry moves closer to the processor, the available space inside the package substrate becomes increasingly limited. At the same time, PCB real estate is becoming more valuable than ever.

One way to resolve this conflict is to integrate capacitors directly into the package. Embedded MLCCs relocate capacitance from the PCB surface into the package substrate itself. This frees valuable board space for additional components while simultaneously reducing the electrical distance between the processor and the capacitor. Shorter current paths further improve power delivery performance and reduce parasitic resistance.

TAIYO YUDEN has pursued this concept for many years. Its embedded MLCCs feature flat copper electrodes, which improve manufacturing yield during substrate fabrication. Wide electrode structures also allow multiple vias to be used simultaneously, further reducing parasitic resistance.

Another advantage is the secure mechanical fixation of the components during resin encapsulation, increasing process reliability during manufacturing. Today, the company has accumulated more than ten years of experience in the mass production of embedded capacitor solutions.

Low-ESL Technology Improves High-Frequency Performance

Besides high capacitance, low parasitic inductance has become equally important.

The faster a processor changes its current consumption, the more its performance depends on the effective inductance of the decoupling capacitors. Excessive inductance limits the capacitor's ability to supply current during extremely fast transient events. For this reason, TAIYO YUDEN complements its High-Density MLCC portfolio with Low-ESL solutions based on three-terminal technology.

Their special terminal geometry significantly reduces parasitic inductance while simultaneously lowering impedance, particularly at high frequencies. As a result, extremely fast load transients can be compensated even more effectively.

A Portfolio for the Next Generation of AI Servers

As data centers increasingly rely on more powerful AI accelerators, higher current densities, and more compact server architectures, TAIYO YUDEN approaches these challenges from a system perspective rather than focusing on individual components.

The portfolio includes high-density MLCCs for a wide range of supply voltages, embedded MLCCs for advanced package architectures, and low-ESL capacitor solutions designed for highly dynamic load conditions.

The portfolio is complemented by high-performance inductors that work together with the capacitors to enable compact and highly efficient power delivery systems.

The focus is therefore not simply on increasing the capacitance of individual components. Rather, success depends on the interaction of every element within the Power Distribution Network.

Only when voltage regulators, inductors, and MLCCs are carefully optimized as an integrated system can the extremely dynamic load profiles of modern AI accelerators be managed reliably.

MLCCs at the Heart of Modern Power Delivery

The power requirements of today's AI processors continue to grow at an extraordinary pace. At the same time, demands for higher energy efficiency, greater power density, and faster transient response are increasing. As a result, the Power Distribution Network is becoming a key differentiating factor for next-generation AI computing platforms.

In addition to MLCCs designed for embedded substrate applications, TAIYO YUDEN offers a lineup of cutting-edge X6S components, which includes 0402-inch (22µF, 47µF), 0603-inch (22µF, 47µF, 100µF), and 0805-inch (47µF, 100µF, 220µF) MLCCs.

Multilayer ceramic capacitors store energy directly adjacent to the processor, stabilize the supply voltage during extremely fast load transients, and thereby enable the performance levels required by modern AI accelerators.

With its High-Density MLCCs, Embedded MLCCs, and Low-ESL technologies, TAIYO YUDEN directly addresses these challenges. By combining advanced material science, precision manufacturing, and decades of expertise in passive components, the company has developed a portfolio specifically designed for the power delivery requirements of future AI data centers.

TAIYO YUDEN therefore positions itself not merely as a component supplier, but as a technology partner for engineers developing the next generation of AI server power architectures.

(ID:50925536)