The high-performance computing market has soared, and the value of insignificant components has increased by 8 times.

---

As the market scale of high-performance computing (HPC) systems, particularly AI servers, continues to expand, the performance and power consumption levels of core processors, including CPUs, GPUs, NPUs, ASICs, FPGAs, and memory and networking chips, are also increasing. Alongside these improvements, power management capabilities have become even more critical due to the growing energy consumption of HPC systems, especially AI servers, which places higher demands on the system and the primary components’ power management abilities.

In AI servers, the CPU, GPU cards, memory (DDR4, DDR5, HBM), and various interfaces all require power. At this point, the power management system plays a crucial role, not only involving AC/DC power supplies and DC/DC converters but also utilizing passive components such as inductors and capacitors, which play a key role. With the improvement in system performance and power consumption, there is an increased demand for both the performance and quantity of these passive components.

High-performance passive components can provide stable voltage and current, ensuring that HPC systems like AI servers run smoothly and maintain fast transient response and low ripple. Low-loss passive components can improve the energy efficiency of AI servers, enhance the efficiency of key components, and promote environmental protection. To ensure the reliability and stability of AI servers, there is a greater need for high-quality inductors.

01 Power Supply Challenges for AI Systems

Compared to ordinary servers, AI servers require higher configurations and consume more energy. Since the power consumption of AI servers is 6 to 8 times higher than that of ordinary servers, the requirements for power supplies are correspondingly higher. Currently, general servers typically require two 800W power supplies, while AI servers may need up to four 1800W power supplies.

As server performance increases, the number of accompanying inductor transformers will inevitably rise. For example, chip inductors: according to one report, due to the increase in the number of GPUs, AI servers require between 24 and 48 inductors. At $1 each, the value of inductors in AI servers is 60% to 220% higher compared to ordinary servers.

Additionally, in AI servers, multi-phase or coupled inductors in integrated forms are gradually replacing single inductors. To address heat dissipation and loss issues, ultra-thin applications and power module power supplies will become more widely used.

Data centers require more and more AI acceleration cards, equipped with large numbers of processors (xPUs). These processors often use large-scale parallel computing solutions. Compared to ordinary CPUs, xPUs have many small cores, which are beneficial for neural network training and AI inference. However, when xPUs perform AI calculations and transmit data, they consume significant amounts of power. In other words, xPUs are very power-hungry chips, and their strict power requirements pose new challenges for AI acceleration cards, which can affect system performance.

When AI systems are working, especially when handling deep learning and inference workloads, they require extremely high computational power. At the system level, AI accelerators play a critical role in providing results that are nearly real-time. All xPUs have multiple high-end cores, consisting of billions of transistors, consuming hundreds of amperes of current. The core voltages of these xPUs have dropped to levels around 1V.

The peak current density required by AI acceleration cards is a heavy burden for any motherboard to handle. The highly dynamic nature of workloads and the extremely high current transients cause very high di/dt and voltage spikes lasting several microseconds, which can be destructive and potentially damaging to xPUs. When the average workload of AI systems persists for a long time, decoupling capacitors may not always provide the energy needed to meet immediate demands. At this point, it is necessary to eliminate transients in AI accelerators to avoid damage to the entire power distribution network.

Currently, the requirements for xPU voltage regulators (VRs) differ significantly from standard PoL voltage regulators. Some applications require supplying over 1000A of current to xPUs at voltages below 1V. At this point, power consumption must be carefully controlled, or the system will struggle to operate stably.

02 Server Power Systems Evolve

The miniaturization of processors has led to lower supply voltages, but the consumed current has not decreased, resulting in continued increases in power consumption. One of the problems caused by the trend towards lower voltages and higher currents is how to improve the system’s rapid response capability to load fluctuations.

As voltage decreases, the allowable tolerance becomes very small. For instance, to prevent processor malfunctions, if the core voltage is provided with ±3% accuracy, the tolerance at 1V must be kept within ±30mV. For server-specific power supplies, even under conditions where the load changes abruptly to over 1000A, the output voltage must remain as stable as possible.

In practical applications, the trend towards lower voltages and higher currents continues, typically addressed through higher switching frequencies and multiphase designs. Switching operations at higher frequencies allow for smaller components (such as capacitors and inductors) to manage and smooth the energy flow in input and output circuits. For converters based on conventional silicon power semiconductor devices, typical switching frequencies range from 30 to 80 kHz. At these frequencies, commonly accepted capacitors can be used cost-effectively. However, above this frequency range, parasitic effects can lead to excessive resistance losses and self-heating.

Although increasing the frequency greatly improves the response to load changes, it also significantly increases the losses in switching elements. Additionally, using large-capacity external capacitors can suppress voltage fluctuations in high-current applications to some extent, but this increases installation space and capacitor costs.

Considering these factors, Trans-Inductor Voltage Regulators (TLVRs) are currently the mainstream circuit configuration solution for addressing rapid load fluctuations in low-voltage, high-current applications. TLVRs connect each phase switch to an inductor with an additional winding, then serially connect the windings of each phase and compensation inductors into a loop, providing current simultaneously to each phase. TLVRs enable processors to achieve high transient response performance, meeting load requirements without a significant drop in supply voltage, while reducing power losses and keeping output capacitance values small, thus reducing installation area and system costs.

03 More Inductor Solutions

In the power management systems of high-performance computing systems, particularly AI servers, there is an increasing use of inductor solutions, including TLVRs, integrated inductors, chip inductors, and ultra-thin integrated inductors.

Chip inductors are responsible for powering the front end of chips, primarily used for voltage and current conversion. They are commonly found in power management ICs (PMICs) and FPGA power supply circuits. In high-performance computing systems, chip inductors, capacitors, MOSFETs, and driver ICs together form the power supply circuit, meeting the power needs of GPUs and CPUs.

Currently, mainstream chip inductors use ferrite materials, but these have poor saturation characteristics. As power modules become smaller and the current increases, the size and saturation characteristics of ferrite inductors have become insufficient to meet the requirements of high-performance GPUs. In recent years, a type of metal soft magnetic material inductor has emerged, offering higher efficiency, smaller size, and better responsiveness to high current changes. Chip inductors made from metal soft magnetic materials can be used at switching frequencies ranging from 500kHz to 10MHz.

There is another type of chip inductor that uses semiconductor thin-film technology and photolithographic processing, differing from traditional wound inductors and integrated inductors. The main feature of semiconductor thin-film technology is the ability to produce chip inductors in full batches, improving production efficiency. Traditional power modules use SIP technology to combine chips and inductors in a single package base, integrating the power inductor and package base into a single unit. Compared to traditional SIP, which requires “chip + inductor + base,” the semiconductor thin-film technology approach integrates the chip, inductor, and other components into a complete power module and peripheral circuit, further reducing the size of the power module, increasing power density, and reducing costs.

This type of chip inductor uses new magnetic materials with good permeability and saturation current. At a frequency of 6MHz, the material loss of the inductor accounts for a low proportion of the total inductor loss.

04 Capacitors Are Also Important

In the power management systems of high-performance computing, alongside inductors, capacitors and thermistors are also undergoing updates and improvements.

Currently, the proportion of AI servers in the overall high-performance computing market is still relatively low, so no market research institutions have yet quantified the consumption of MLCCs (multilayer ceramic chip capacitors) in AI servers. However, based on development trends, passive component distributors generally have a positive outlook on capacitors, especially MLCCs, in AI servers. There is expected to be a noticeable growth trend in the second half of 2024, with specifications and prices of MLCCs significantly increasing.

From a technical perspective, all processors in computing systems require capacitors to function. Traditionally, these capacitors were tantalum or polymer capacitors. To reduce reliance on decoupling capacitors, a small number of Class II MLCCs (such as X5R, X6S, or X7R devices) can be placed directly near the processor. Currently, some manufacturers are working on embedding aluminum polymer decoupling capacitors within the chip carrier in the package, working together with on-chip silicon capacitors. This can overcome the decoupling challenges faced by high-performance processors and support higher converter frequencies, potentially up to 10MHz in the future.

05 Opportunities for Passive Component Manufacturers

At NVIDIA’s recent GTC conference, Delta Electronics, a major server OEM, indicated that in the power conversion systems of AI servers, maintaining a stable voltage of 0.8V under rapidly surging currents is a critical role played by inductors. They must be able to operate stably under high currents and low voltages.

AI servers equipped with NVIDIA’s new Blackwell architecture accelerator chips consume 1000W to 1200W of power. The number of inductors used is 2 to 3 times higher than in general servers, and because the power consumption has significantly increased, the required specifications of inductors are higher, making their average selling price (ASP) 5 to 8 times higher than that of general servers. Additionally, as the penetration rate of DDR5 increases, more and better inductors are required.

The power consumption of AI servers has significantly increased. To improve transient response performance, TLVR inductors are added, with 5 to 10 per AI server, and the unit price of TLVR inductors is 3 to 5 times higher than that of regular inductors.

Not just the latest AI servers, but an increasing number of high-performance computing systems require more and better inductors. When upgrading CPUs in general servers, the number of inductors significantly increases. For example, when upgrading from Eagle Stream to Birch Stream, since the CPU power consumption increases by about 50%, the number of inductors needs to increase by 50% to 70%.

For major passive component manufacturers, especially those producing high-quality inductors, new business opportunities are emerging. Leading companies in the industry include TDK, Yageo, Sunlord Electronics, Taigene Technology, ITG, and Eaton.

As mentioned earlier, the number of chip inductors used in the power management systems of high-performance computing is increasing. This is not only good news for international giants but also presents an opportunity for Chinese domestic enterprises to improve product quality and market share. China’s chip inductor industry started late and initially lagged behind international giants, particularly TDK, Murata, Chilisin, and TAIYO Yuden. In recent years, China’s Sunlord Electronics has risen to the top five globally. Other notable domestic chip inductor companies include Bokai New Materials, Magtech Technology, Yitong New Materials, Tian Tong Shares, Dongmu Shares, and Hengdian Magnetics.

06 Conclusion

As the market scale of high-performance computing systems, particularly AI servers, continues to expand, the requirements for key chip components are becoming increasingly stringent. Not only do high-performance processors like GPUs and CPUs face these demands, but power management systems and related chips and components are also seeing significant increases in both quantity and quality requirements.

For inductors and capacitors, which are unassuming but indispensable components of power management systems and used in large quantities, the ever-increasing power consumption of computing systems provides a stage for them to fully demonstrate their performance and utility. New technologies and materials are expected to emerge continuously.

For passive component manufacturers, international giants with high-quality products will continue to enjoy better business opportunities. For Chinese domestic companies, the vast domestic market provides ample space for them to grow, offering more opportunities to capture market share from international giants.