GTC 2025 Preview: Chinese Firm Foresees Focus on Power Management, Cooling, Communication & PCBs
24 March 2025 路 Uncategorized 路
Source: 路 https://finance.technews.tw/2025/03/15/gtc-2025-forecast-report/
NVIDIA鈥檚 GPU Technology Conference (GTC) in 2025 is approaching, and according to a prediction by China's Guosen Securities, this year鈥檚 conference will emphasize power management, liquid cooling solutions, communication technologies, and high-density printed circuit board modules (PCBs). These advancements are projected to significantly enhance the efficiency of AI computations and improve data center infrastructure performance.
Guosen Securities anticipates that High-Voltage Direct Current (HVDC) technology is becoming increasingly prevalent due to rapidly rising server power demands driven by artificial intelligence applications. At this year鈥檚 GTC, Delta Electronics and Lite-On Technology are expected to introduce 400V/800V HVDC products designed to improve data center efficiency. Furthermore, supercapacitors and lithium battery backup power units have gained attention for their ability to manage high instantaneous power demands in AI computations while bolstering server stability.
Liquid cooling solutions are anticipated to become standard practice within AI servers. NVIDIA鈥檚 forthcoming GB300 server is expected to feature a B300 GPU with a thermal design power (TDP) of 1400W and will fully adopt cold plate-based liquid cooling technologies, potentially reducing energy consumption by over 30% compared to air cooling鈥攖hereby increasing data center computing density. While current solutions primarily utilize cold plates, the report suggests that immersion technology represents a likely future direction for server cooling.
Regarding AI server communication technology, industry experts predict NVIDIA will launch an 115.2Tbps co-packaged optics (CPO) switch at this year鈥檚 GTC conference. This high-bandwidth interconnect is expected to reduce data transmission power consumption and costs while improving the efficiency of AI computing clusters, potentially driving upgrades in cloud data centers and AI training environments.
PCBs are evolving alongside advancements in server architecture; Guosen Securities predicts continued demand for High-Density Interconnector (HDI) boards and high-layer count circuit boards. The internal structure of GB300 servers is anticipated to revert to an HGX UBB+OAM design, enhancing reliability and cooling efficiency within computing modules. Furthermore, PTFE materials and HVLP5 copper foil technology are emerging as key development directions in PCB manufacturing.
The report also suggests potential unveilings at this year鈥檚 event, including NVL288 racks and the next-generation Rubin architecture鈥攚here four NVL72 cabinets will be interconnected via back-end cables to form an NVL288 rack. CPO and immersion liquid cooling are expected to become standard features in the Rubin design.
(Lead image source: Nvidia)
Guosen Securities anticipates that High-Voltage Direct Current (HVDC) technology is becoming increasingly prevalent due to rapidly rising server power demands driven by artificial intelligence applications. At this year鈥檚 GTC, Delta Electronics and Lite-On Technology are expected to introduce 400V/800V HVDC products designed to improve data center efficiency. Furthermore, supercapacitors and lithium battery backup power units have gained attention for their ability to manage high instantaneous power demands in AI computations while bolstering server stability.
Liquid cooling solutions are anticipated to become standard practice within AI servers. NVIDIA鈥檚 forthcoming GB300 server is expected to feature a B300 GPU with a thermal design power (TDP) of 1400W and will fully adopt cold plate-based liquid cooling technologies, potentially reducing energy consumption by over 30% compared to air cooling鈥攖hereby increasing data center computing density. While current solutions primarily utilize cold plates, the report suggests that immersion technology represents a likely future direction for server cooling.
Regarding AI server communication technology, industry experts predict NVIDIA will launch an 115.2Tbps co-packaged optics (CPO) switch at this year鈥檚 GTC conference. This high-bandwidth interconnect is expected to reduce data transmission power consumption and costs while improving the efficiency of AI computing clusters, potentially driving upgrades in cloud data centers and AI training environments.
PCBs are evolving alongside advancements in server architecture; Guosen Securities predicts continued demand for High-Density Interconnector (HDI) boards and high-layer count circuit boards. The internal structure of GB300 servers is anticipated to revert to an HGX UBB+OAM design, enhancing reliability and cooling efficiency within computing modules. Furthermore, PTFE materials and HVLP5 copper foil technology are emerging as key development directions in PCB manufacturing.
The report also suggests potential unveilings at this year鈥檚 event, including NVL288 racks and the next-generation Rubin architecture鈥攚here four NVL72 cabinets will be interconnected via back-end cables to form an NVL288 rack. CPO and immersion liquid cooling are expected to become standard features in the Rubin design.
(Lead image source: Nvidia)