The increase in computing power is driving rack power density ever higher. The single-rack power of NVIDIA's GB200 NVL72 is 120kW, and that of the GB300 NVL72 is 132kW. In the future, for the VR200 NVL series based on the Rubin architecture and the VR300 NVL series based on the Rubin Ultra architecture, the single-rack power will reach hundreds of kilowatts or even 1MW.
China is advancing the domestic substitution of chips, giving rise to a series of technology companies such as Huawei, Cambricon, and Moore Threads. Compared with NVIDIA, domestic chips have relatively backward process nodes and relatively low single-card computing power, so breakthroughs will be sought in super nodes and clusters. The single-rack power of Huawei's Ascend 384 super node has reached 42kW. In the future, with the iteration of chip technology, the single-rack power will be even higher.
Figure Trend of China's AI Computing Power Development
The increase in rack power density brings many problems to data centers, mainly manifested in the following aspects:
1. Layout Issues of Power Supplies within the Rack
After the rack power increases, the current increases while the voltage remains unchanged. Taking a 1000kW rack as an example, if three-phase AC380V is adopted, the current per phase is 1515A. Whether using PDUs or copper busbars to supply power to servers, there is not enough space for layout within the rack. If rack-mounted PSUs are used, there will be no space left in the rack to place servers.
2. Proportion Issues of Gray and White Areas
Taking a 20MW data center as an example, if the ratio of the gray area to the white area is 1:1 when the power per rack is 15kW, then when the power per rack increases to 60kW, assuming the power supply and cooling solutions remain unchanged, the ratio of the gray area to the white area may change to 4:1. Extrapolating from this, if the rack power density continues to increase while the power supply and cooling solutions of the data center remain unchanged, the proportion of the gray area will become higher and higher, and the ratio will be seriously unbalanced.
3. Bottleneck Issues in Improving Power Supply Efficiency
The architecture of traditional power supply systems has been basically determined, and the efficiency of various equipment is already close to the limit. After the rack power density increases, the transmission current is larger, resulting in greater line losses and more heat generated. This not only affects the power supply efficiency of the data center but also brings more pressure to the cooling system.
To avoid the above problems, increasing the power supply voltage of the data center is an effective approach. As the rack power density increases, the power supply voltage level is gradually raised. After the voltage is increased, the current becomes smaller, and issues such as power supply layout and transmission losses within the rack will be improved. At the same time, the voltage increase provides an opportunity to enhance the power density of power supply equipment and reduce equipment nodes, thereby saving floor space and further improving the ratio of gray to white areas in the data center.
There are two ways to increase the power supply voltage of data centers: one is alternating current (AC), and the other is direct current (DC).
The AC approach is already being explored overseas, increasing the current three-phase AC480V to three-phase AC600V, still implemented by UPS. Since AC600V is not much different from AC480V, many of the original power devices can still be used in common, and both are within the low-voltage AC range without violating relevant specifications, making it relatively easy to implement. However, because the voltage increase is not significant, it is only applicable to situations where the rack power density does not increase much. Meanwhile, AI users led by NVIDIA have explicitly proposed the introduction of DC power supply and the popularization of DC800V applications by 2027. Therefore, it is predicted that the future overseas AI market will be dominated by DC800V power supply, while AC600V will be used as a supplementary solution for DC power supply in some retrofitted or small and medium-sized data centers with relatively low power density.
For the domestic market, the leap from AC380V to AC600V or AC690V is somewhat large, and modifications based on existing UPS are relatively complex. At the same time, its voltage increase has certain limitations; above AC690V, it will enter the medium-voltage area, so the possibility of its adoption is very small. In contrast, voltage increase based on DC technology will be much easier, with mature technology and equipment foundations. Therefore, DC800V or ±400V will be the future development trend, where ±400V can be used as DC800V or as DC400V.
Figure Matching Relationship between Rack Power and Power Supply Voltage
Excerpted from the "Data Center 800V DC Power Supply Technology White Paper" (2.0)
Recently, the China Engineering Construction Standardization Association (CECS) officially issued the "Second Batch of Association Standard Formulation (Revision) Plan for 2025". In this plan, the "Technical Specification for DC Power Supply Systems in Data Centers", which is under the jurisdiction of the Data Center Professional Committee, has been successfully included in the plan, and the inaugural meeting of the compilation group and the first working meeting were held in Beijing on January 22. This standard is edited by Zhongshu Wisdom (Beijing) Information Technology Research Institute Co., Ltd. and China Information Technology Designing & Consulting Institute (CITDCI), and the main participating compilation units include: Douyin Vision Co., Ltd., Beijing Kuaishou Technology Co., Ltd., Shenzhen Tencent Computer Systems Company Limited, Alibaba Cloud Computing Co., Ltd., Beijing VNET Broadband Data Center Co., Ltd., GLP Pura Data Technology (Shanghai) Co., Ltd., Shiyuan Technology Engineering Co., Ltd., Huaxin Consulting Design Research Institute Co., Ltd., Delta Electronics (China) Co., Ltd., Shenyang Microcon Flywheel Technology Co., Ltd., Hefei Sunshine Source Intelligent Technology Co., Ltd., Nanjing Borland Electronic Technology Co., Ltd., Schneider Electric (China) Co., Ltd., etc.
This standard is applicable to the design, construction, and acceptance of DC power supply systems in newly built, reconstructed, and expanded data centers. We sincerely invite organizations with outstanding practical experience in the field of data center power supply and distribution technology to participate jointly and work together to create benchmark achievements that lead industry development.
Contact: Luo Yuxi, 13716595411 (same as WeChat)