2023-10-14-施耐德电气-2023人工智能带来的颠覆_20页_1mb

AI and machine learning (AI) are rapidly transforming data centers, driving higher power density requirements that challenge traditional infrastructure. This white paper highlights key growth statistics, such as AI power demand reaching 13.5-20 GW by 2028, compared to 54 GW for all data centers. Training and inference workloads differ in power density, with training typically operating at near-peak utilization due to high parallel processing needs, often requiring densities up to 100 kW per cabinet. Challenges emerge in power distribution, cooling, cabinet design, and software management. For example, standard voltages like 120/208V are inadequate, necessitating upgrades to 240/415V for efficient power delivery. Cooling must transition from air-based systems to liquid cooling to handle high thermal design power (TDP), with risks like leakage addressed through advanced designs. Cabinet modifications are required, including increased width, depth, height, and weight capacity to fit denser servers. Software tools like DCIM and EPMS support resource monitoring and management, reducing risks. Future trends include higher-density components, liquid cooling enhancements, and integrated grid interactions.

Overall, AI adoption demands robust infrastructure redesign, with careful planning for both new and existing facilities to ensure reliability and efficiency in the face of increasing computational loads.