What kind of server power does AI need?
Produced by ZhiNeng Tech
The rapid advancement of AI technology has triggered a huge demand for computing power, highlighting the critical role of AI server power supplies.
This article will analyze the current development of AI server power supplies from multiple dimensions, including their importance, technical characteristics, market demand, and industry competition.
01
Demand Trends for AI Server Power Supplies
The advancement of AI technology has led to a surge in power consumption. According to the Artificial Intelligence Index Report 2023, the single training session power consumption of advanced AI models such as DeepMind Gopher, Hugging Face BLOOM, OpenAI GPT-3, and Meta OPT reached 1066MWh, 433MWh, 1287MWh, and 324MWh, respectively.
Data from the IEA indicates that the power consumption of NVIDIA chips in 2023 reached 7.3TWh, and it is projected to increase tenfold by 2026. In the context of growing power demand, high-efficiency server power supplies have become crucial. Compared to the 96% conversion efficiency of traditional server power supplies, AI server power supplies need to meet a higher standard of 97.5%-98% to reduce energy loss and support higher power density GPUs.
The iteration of AI models and GPU technology has driven continuous improvement in the power density of AI server power supplies. Taking the H100 server rack as an example, a 10.2kW power requires a 19.8kW power supply; the 120kW power of the NVL72 rack demands a 198kW power module. The power density of AI server power supplies can reach 100W/cubic inch, far exceeding the 50W/cubic inch of ordinary server power supplies.
02
Technical Architecture and Features of AI Server Power Supplies
The architecture of AI server power supplies consists of three layers: UPS, AC/DC, and DC/DC, each playing a role in ensuring the continuity and stability of power supply.
● UPS: Provides temporary power during grid interruptions to protect data from loss.
● AC/DC: Converts grid AC power to 50V DC power suitable for server use.
● DC/DC: Further reduces the 50V DC power to 0.8V acceptable for chips.
The AC/DC power module includes the power supply module (PSU) and the power management controller (PMC). For example, the NVL72 server rack requires six power racks, totaling 36 modules with a combined power of 198kW. The DC/DC module needs to be tightly integrated with the load to reduce power network loss.
The target power density for AI server power supplies is 180W/cubic inch, posing new challenges for cooling technology. Liquid cooling, which effectively improves temperature control, is gradually replacing traditional air cooling solutions.
By 2024/25, the demand for AC/DC and DC/DC in AI server power supplies is expected to reach 10.2GW/11.8GW and 3GW/4GW, respectively. The market size is projected to expand to 15.6 billion to 27.2 billion yuan by 2025. Currently, Taiwanese manufacturers lead the market, but domestic brands are catching up and are expected to secure a place in the future through continuous R&D investment and technological accumulation.
Summary
As the cornerstone of AI technology development, the technological innovation and market expansion of AI server power supplies will accompany the continued prosperity of the AI field. Facing the future, domestic brands need to seize opportunities, meet challenges, and jointly promote the progress and innovation of the AI server power supply industry.
The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.
