Applications of multilayer ceramic capacitors and polymer tantalum capacitors in AI servers

AI servers are a key growth market 

For decades, traditional servers have been the cornerstone of data centers. They are designed to handle various workloads, including hosting websites, managing databases, and running enterprise applications. The architecture of traditional servers typically includes a combination of CPUs (central processing units), RAM (random access memory), storage, and networking. 

Traditional servers heavily rely on CPUs, which are versatile and capable of handling multiple tasks simultaneously. RAM is essential for fast data access and smooth application performance, while storage consists of a combination of HDDs (hard disk drives) and SSDs (solid-state drives) for data storage. Networking provides high-speed interfaces for data transfer within and outside the server. Traditional servers play a vital role in data storage and processing across various fields. 

Today, AI servers have become a hot development trend. AI servers are specifically optimized for artificial intelligence and machine learning workloads. These servers are equipped with specialized hardware and software to accelerate AI tasks, such as training deep learning models and performing complex data analysis. Key components of AI servers include GPUs (graphics processing units), TPUs (tensor processing units), NVMe (non-volatile memory express) storage, and high-bandwidth memory (HBM). 

Unlike traditional servers, AI servers rely heavily on GPUs. GPUs can process multiple parallel tasks simultaneously, making them ideal for AI and machine learning workloads. TPUs, developed by Google, are custom designed to accelerate machine learning tasks. Compared to GPUs, they offer higher performance per watt for specific AI workloads. NVMe drives provide significantly faster read and write speeds, which are crucial for handling large datasets common in AI applications. HBM is an enhanced memory architecture that provides faster data access speeds, reducing bottlenecks in AI processing. 

In AI servers, SSDs (solid-state drives) are used for data storage in AI and power-loss protection. Polymer tantalum capacitors can be used for hold-up in enterprise SSDs. During power outages, polymer tantalum capacitors release energy to allow DRAM to write data back to NAND flash memory, making them critical components paired with enterprise SSDs. 

MLCCs meet the power consumption and power stability requirements of AI servers 

Multilayer Ceramic Capacitors (MLCCs) are widely used in electronic products, but their role in AI servers differs significantly from that in traditional servers or general electronic devices. These differences mainly stem from power consumption characteristics, processor architecture, and the greatly increased demand for power supply stability. 

AI servers utilize a large number of GPUs and AI accelerator cards (such as NVIDIA H100, AMD MI300, etc.), with individual card power consumption reaching several hundred watts or even exceeding 1 kW. The power load changes rapidly, and there are extremely high requirements for signal integrity (SI) and power integrity (PI). The core function of MLCCs in AI servers includes power decoupling to stabilize the power supply voltage to processors/GPUs and suppress voltage drops caused by instantaneous load fluctuations. Their high-frequency response must be faster, requiring arrays of low-ESL (Equivalent Series Inductance) MLCCs placed close to the chip packaging. 

In terms of filtering functionality, it is necessary to filter out high-frequency noise in the power supply and signal paths to ensure signal stability for AI accelerators and high-speed interfaces (PCIe Gen5, CXL, HBM). The frequency range is broader (from hundreds of kHz up to the GHz level), requiring a combination of various packaging and dielectric types. Regarding bulk capacitance, it must provide current during sudden load changes to reduce the impact of VRM response delay on voltage. This increases the capacitance demand, and capacitors such as tantalum and polymer types must be used in combination to balance low ESR and large capacity. Additionally, these capacitors can be used to suppress EMI/EMC, reducing electromagnetic interference generated by AI computations and high-speed communications. This requires high-Q, low-loss dielectrics and layout optimization tailored for high-speed transmission lines. 

Compared to PC servers, communication equipment, or industrial control systems, AI servers have different MLCC requirements, including power consumption variation speed. Traditional servers experience CPU/GPU power switching changes on the order of microseconds to milliseconds, whereas AI servers’ GPU/TPU instantaneous load changes occur as fast as nanoseconds to microseconds, necessitating ultra-high-speed decoupling. In terms of capacitance configuration, traditional servers mainly use capacitance values in the microfarad to several hundred microfarads range, distributed around the motherboard and VRM. AI servers require larger capacitance values and must densely place low-ESL MLCCs near the package (even in 0201/01005 sizes). 

In terms of frequency response, traditional servers primarily operate at the MHz level, whereas AI servers must cover broadband noise suppression ranging from several hundred kHz to several GHz. Regarding temperature and lifespan requirements, traditional servers typically operate within a commercial temperature range of 0~85°C or 105°C, while AI servers demand high reliability, running continuously in environments between 85~125°C and requiring resistance to DC bias effects.

Additionally, in packaging and layout, traditional servers use standard sizes (0402~1210), whereas AI servers tend toward smaller packages (01005/0201) and high-capacity multilayer structures, combined with embedded PCB MLCC technology. For reliability standards, ordinary IEC/JEDEC grades suffice for traditional servers, but AI servers must meet AEC-Q200 or data center-level durability and vibration resistance specifications.

Due to the higher energy demands of AI and big data applications, better electromagnetic interference (EMI) and noise resistance are required. Therefore, MLCCs need more electrode layers, thinner dielectrics, larger effective areas, and higher reliability to achieve higher capacitance per unit volume and higher effective capacitance. The materials used are also shifting from X5R to X6S. High-capacitance-density MLCCs have become key components in AI server applications.

Typical AI server MLCC configurations mostly use low ESL arrays, paralleling multiple 0201/01005 MLCCs placed close to the GPU/TPU package to provide response capabilities on the order of 1 nanosecond. They also employ mixed capacitor networks, combining MLCCs, polymer capacitors, and tantalum capacitors to form broadband filtering and energy storage structures. Additionally, high-temperature and high-reliability specifications are required, such as using X7R, X8R, or C0G/NPO materials, to ensure performance stability under high temperatures and long-term operation. Using embedded capacitors is another strategy; some advanced AI motherboards integrate MLCCs directly into the PCB layers to shorten parasitic inductance paths. Future AI servers will require more MLCCs than ever before, but for cost considerations, MLCCs will replace tantalum capacitors in accelerators.

From a materials technology perspective, MLCCs used in AI servers need better inner electrode continuity, uniform BT grains, and a homogeneous grain structure. The BT core can provide a coverage layer, while the shell offers insulation resistance (IR) to achieve high-capacitance products. Smaller and more uniform BT grains and a homogeneous core-shell structure represent more advanced designs.

Polymer tantalum capacitors offer advantages in high-power, high-transient environments of AI servers

Polymer tantalum capacitors are primarily used in AI servers as large-capacity, low-ESR mid-to-high-frequency energy storage and decoupling components, working alongside MLCCs to form a broadband power stability network. Compared to traditional tantalum or aluminum electrolytic capacitors, they offer significant advantages in the high-power, high-transient environments of AI servers.

Polymer tantalum capacitors feature extremely low ESR (equivalent series resistance), with polymer electrolyte ESR as low as 5~20 mΩ—an order of magnitude lower than traditional tantalum capacitors. This allows them to quickly supply power during sudden GPU/TPU load changes, reducing voltage drops. They also exhibit high ripple current tolerance, handling higher ripple currents than liquid electrolytics, making them suitable for high-frequency power switching and rapid pulse loads in AI servers.

Additionally, polymer tantalum capacitors offer excellent temperature stability, with minimal electrical performance variations across -55°C to +125°C, ensuring stability in data centers under prolonged full-load operation. Their long lifespan and high reliability, with solid polymers eliminating drying-out issues, make them ideal for 24/7 high-load AI data center demands.

Polymer tantalum capacitors also feature low noise, with superior high-frequency performance compared to liquid electrolytics, ensuring signal integrity for high-speed interfaces (PCIe Gen5, CXL, HBM). They eliminate the risk of "short-circuit failure explosions," as the polymer conductive layer typically fails open under stress, enhancing system safety and reducing fire risks.

Polymer tantalum capacitors are commonly used in AI servers at the motherboard VRM output, GPU/TPU power supply modules, and HBM memory power supply areas. Configuring polymer tantalum capacitors can compensate for the insufficient capacitance of MLCCs by providing large current energy storage capabilities. In large-capacity mid-frequency filtering applications, MLCCs have high capacitance and good high-frequency response but lack low-frequency energy storage; polymer tantalum capacitors supplement capacitance in the μF to mF range but require selecting a high rated voltage (matching the VRM output voltage margin).

Additionally, polymer tantalum capacitors can be used to optimize power supply transient response. When an AI accelerator card transitions from idle to full load, the instantaneous current rises sharply (high di/dt). At this time, paralleling low-ESR polymer tantalum capacitors helps reduce transient voltage drops. For stable power rail applications, data center-grade AI chips have very low tolerance for power supply voltage fluctuations (±2% or even less), so it is necessary to combine MLCC arrays with polymer tantalum capacitors to achieve wide-frequency decoupling.

Polymer tantalum capacitors also adapt well to high-temperature, high-load environments. The ambient temperature inside server racks can exceed 85°C, so polymer tantalum capacitors rated for 105°C to 125°C must be selected. Regarding reliability, since downtime costs for AI servers are high, military-grade or automotive-grade (AEC-Q200 or higher) polymer tantalum capacitors should be used.

In AI server power supply design, polymer tantalum capacitors and MLCCs have a complementary relationship. MLCCs handle high-frequency decoupling (MHz to GHz), offering fast response but limited capacitance, while polymer tantalum capacitors manage mid- to low-frequency energy storage and decoupling (kHz to MHz), providing large capacitance and low ESR. The common approach combines multiple low-ESL MLCCs with several high-capacitance polymer tantalum capacitors in parallel, creating a stable power network that covers the full frequency range from kHz to GHz.

High-performance products meeting AI server demands

For AI server applications, YAGEO has launched high-capacitance MLCCs--the HC series, offering X5R, X6S, and X7R high-capacitance MLCCs with capacitance ranges from 1 µF to 100 µF, case sizes from 0201 to 1210, and voltage support up to 4 V~50 V. Higher-grade X7R MLCCs come in 0402~2220 packages, supporting 4 V~100 V voltages and capacitance ranges of 1 µF~47 µF. These high-capacitance MLCCs are designed to meet the electronics industry's demands for miniaturization, higher voltages, and higher frequencies, providing highly reliable MLCCs to support big data processing needs with higher capacitance stability, low ESR, and enhanced reliability.

The high-capacitance HC series X5R and X6S MLCCs feature higher capacitance, small and thin profiles, continuous operation in harsh environments, low ESR, energy efficiency, better VCC (voltage coefficient of capacitance), high-reliability tolerance, and robust terminal metals, making them ideal for AI server requirements.

In the polymer tantalum capacitor product line, YAGEO offers the A700/A720/A798 aluminum polymer series for AI servers. These capacitors use surface-mount technology, operate in temperatures from -55°C to +105°C/125°C, support voltages from 2 V to 35 V, and offer capacitance ranges from 6.8 µF to 680 µF with ultra-low ESR (3 mΩ to 70 mΩ). They feature non-ignition failure modes, solid counter electrodes (no drying out), low capacitance loss at high frequencies, no voltage derating, 125°C tolerance, low DC leakage, polymer cathode technology, 100% accelerated steady-state aging, 100% surge current testing, self-healing mechanisms, and EIA-standard case sizes.

Conclusion

In summary, MLCCs and polymer tantalum capacitors each play irreplaceable roles in AI server power design. The former ensures stable power supply for high-speed computing cores under transient loads with ultra-low ESL and excellent high-frequency decoupling performance, while the latter provides robust energy buffering and long-term reliability with large capacitance, low ESR, and superior mid-to-low-frequency filtering. Through their complementary collaboration across frequency bands and load characteristics, AI servers can simultaneously meet the demands of high-performance computing, stringent power integrity, and long-term stable operation. YAGEO's new MLCCs and polymer tantalum capacitors will undoubtedly play an even more critical role in the future development of AI infrastructure.