Currently, the AI server market is experiencing vigorous growth; however, a series of challenges have emerged, with the surge in power consumption being particularly prominent. According to conservative estimates, by 2030, global data center power consumption will account for one-tenth of the total electricity usage that year. Taking the search scenario as an example, AI search consumes up to 10 times more power than a conventional search engine when processing the same query. Given limited power generation capacity, relying solely on increasing power output to meet such massive electricity demand is not a sustainable solution; reducing the energy consumption of AI devices has become critical.
Power Consumption Challenges
Looking at the trend of server power consumption, the growth rate of AI server power consumption is astonishing. Tokutake Mizuhara, General Manager of the Technology Center at ROHM Semiconductor (Shanghai) Co., Ltd., Shenzhen Branch, pointed out that this is closely related to the development of GPUs.
Taking NVIDIA products as an example, both the rise in power consumption and the pace of product iteration are staggering. The H100 series launched in 2023 has a power consumption of 700W. The B300 launched this year sees its power consumption rise to 1.4kW, while the VR300 series, expected to be launched in 2027, will reach a power consumption of 3.6kW. Meanwhile, NVIDIA’s product iteration cycle has shortened from one series every two years to one series per year.
The high power consumption of NVIDIA GPUs has brought about a series of thorny issues. Current systems are mostly low-voltage systems ranging from 48V to 54V. As the GPU wattage increases, the output current must inevitably rise to meet the power demand. This leads to higher internal losses in cables and greater heat generation, thereby requiring a large amount of copper wire. Taking a 1-megawatt AI server as an example, it requires approximately 4 to 5 tons of copper wire. This not only increases costs but also adds weight and expands the footprint of data centers, while posing severe challenges to heat dissipation and scalability. Furthermore, power consumption accounts for as much as 60% to 65% of server operating costs, making high power consumption a critical factor restricting the development of AI servers.
Path to Breakthrough
To address the aforementioned issues, AI servers are beginning to explore high-voltage systems. Tokutake Mizuhara stated that there are currently two main directions in the market. The first is the ±400V system promoted by Microsoft, Google, and Meta through the Open Compute Project (OCP) alliance; the second is NVIDIA's 800V system. In the Chinese market, both 400V and 800V systems are under development, and there is even discussion about the possibility of upgrading to a 1500V system.
The 800V system offers significant advantages. In terms of scalability, current server DC systems max out at 100kW, whereas the 800V system can achieve over 1MW. In terms of efficiency, compared to the 54V system, the end-to-end efficiency improves by up to 5%, effectively achieving energy savings. Meanwhile, the current in the power bus is reduced, saving copper usage and significantly decreasing thermal losses. Currently, companies such as Delta Electronics are actively developing 800V systems.
Shifting to high-voltage systems will also bring about major changes in the server rack structure. In the current architecture model, AI servers receive power from the power grid, and the power distribution voltage enters the server frame. The frame is divided into a Power Supply Unit (PSU) and an IT unit, responsible for power supply and computing respectively. A Battery Backup Unit (BBU) and Capacitor Backup Unit (CBU) are set in the middle, with the current bus voltage being a 48V to 54V system. The next-generation AI servers will upgrade to high-voltage systems. The power grid remains basically unchanged, while the power supply section is modified by placing the backup unit inside the power supply, and the bus voltage is raised to HVDC, namely ±400V, 800V, or even 1500V.
Opportunities for Manufacturers
Under the high-voltage trend, improving the efficiency of AI servers has become critical. This requires increasing AC/DC and DC/DC efficiency on the power supply side and boosting power density on the server rack side.
Tokutake Mizuhara stated that this poses three requirements for semiconductor manufacturers. First, developing high-efficiency, high-power-density power semiconductor devices requires semiconductor manufacturers to invest substantial R&D efforts. Second, introducing high-voltage technologies, such as 800V or ±400V systems, necessitates joint collaboration and promotion among semiconductor manufacturers, GPU manufacturers, and AI server manufacturers. Third, during mass production, establishing a globalized QCDS (Quality, Cost, Delivery, Service) system is essential to ensure product quality, cost, delivery, and service.
ROHM sees broad opportunities in this. Tokutake Mizuhara stated that as one of the few companies globally possessing both power and analog technologies, ROHM's greatest strength lies in the combination of power devices and analog technology. Its products include power devices based on three materials: Si, SiC, and GaN.
For the 800V system, on the power supply side, ROHM provides a series of SiC products. On the server side, the DC/DC is divided into a primary side and a secondary side. For the primary side, ROHM offers 1200V SiC products, while for the secondary side, it provides various Si products and GaN solutions. This part mainly involves low-voltage DC/DC products with a voltage range of 80V to 150V, requiring lower switching losses and lower on-resistance. ROHM's related products can achieve a power density of 129W per cubic foot. If higher power density is pursued, using GaN can double the power density to 246W per cubic foot.
Tokutake Mizuhara emphasized that using SiC products is highly beneficial for improving the efficiency of power supplies and AI servers. Although the most capable Ruby model currently reaches 98% efficiency, hitting 99% requires SiC, which is also the reason for the widespread adoption of SiC in the AI market.
Currently, ROHM's SiC chips have evolved to the 4th generation. Samples of the 5th generation have begun shipping, with mass production expected to start next year. The 6th generation will be launched in two years, followed by the 7th generation. Compared to the 4th generation, the conduction losses and switching losses of ROHM's 5th-generation SiC chips are significantly reduced, with total losses decreased by approximately 30%, making them more energy-efficient.
Considering application flexibility, ROHM has also formulated various strategies in terms of packaging. Common packages for existing Si or SiC products include TO-220 and TO-247. To facilitate design and miniaturization, ROHM has introduced a new DOT-247 package that combines two TO-247 packages together. If six TO-247 packages are needed, ROHM also offers a compact module, the HSDIP20. The company provides both surface-mount and through-hole products.
To address customer pain points including difficult product development, design challenges, and unstable supply stemming from single-source components, ROHM has collaborated with Infineon. For example, for ROHM's original DOT-247 packaged products, Infineon has also manufactured products with the same package. Conversely, Infineon's entire series of ultra-thin D-DPAK products, which previously lacked a second source, has been entrusted to ROHM for production. This cooperative approach enables both companies to provide customers with more choices and supply guarantees.
Another opportunity product brought by high voltage is the hot-swap device. Hot-swapping refers to removing a slot module while the AI server is running. This operation can cause instantaneous inrush current, which may damage downstream equipment. Therefore, a critical MOSFET is required, along with a current sense resistor and a hot-swap controller.
ROHM's representative Si MOSFET product, "RY7P250BM," has been certified as a recommended device by global cloud platform enterprises. Tokutake Mizuhara emphasized two advantages of their products: first, a wide Safe Operating Area (SOA) range, which facilitates customer usage and design; second, the ability to deliver high current by lowering on-resistance. Additionally, ROHM is developing control ICs for hot-swap applications.
Conclusion
Major architectural transformations often require collaborative efforts across the entire industry. Based on this, ROHM has engaged in deep cooperation with numerous power equipment manufacturers. For example, ROHM has formed a strategic partnership with Delta Electronics, and its EcoGaN™ products are adopted in the 45W output AC adapter "C4 Duo" under Delta Electronics' Innergie brand. Murata's AI servers have also selected ROHM's EcoSiC™. Furthermore, ROHM is deeply involved in Cosel's 3.5kW output AC/DC power supply units. Centered around its own product layout, ROHM is providing practical and feasible strategies to overcome the power consumption challenges in AI data centers.