日本語 
Hitachi, Ltd. has developed a GPU cooling cold plate that reduces the power consumption of data centers, which is increasing with the spread of generative AI. In AI data centers, heat generated from GPUs in servers is recovered by a coolant in a cold plate and released to the outside air via pumps, chillers, cooling towers, etc. As GPUs generate more heat and the number of installed GPUs increases, the load on the cooling equipment used to release the recovered heat to the outside also increases, and power consumption increases as well.
The cold plate prototype we developed employs a unique “U-shaped tunnel channel” structure that directs the coolant towards the heat-generating parts of the GPU, in contrast to conventional structures that flow the coolant along the GPU surface . This structure allows for the rapid recovery of a large amount of heat from the vicinity of the GPU by arranging thousands of fine channels in parallel. After absorbing heat at the heat-generating parts, the coolant folds back in a U-shape and flows to the outlet, preventing the cold and warm coolant from mixing and maintaining consistently high cooling performance. As a result, the GPU can be efficiently cooled even with coolant at a higher temperature than before, and by utilizing the temperature difference with the ambient air, heat can be efficiently dissipated, leading to an expected 30% reduction in the power required for cooling.
Moving forward, Hitachi will use this newly developed cold plate as a starting point and combine it with integrated control of cooling systems to further develop it as one of the cooling technologies that will bring out the full potential of its digital service, “HMAX Data Center. By improving the efficiency and stability of data centers, Hitachi will contribute to the construction of a sustainable digital social infrastructure.
Also Read: Adobe and NVIDIA Expand AI Content Creation Partnership
With the rapid expansion of generative AI, the number of data centers equipped with a large number of GPUs is increasing, and GPU cooling has become a critical factor affecting operational quality and power efficiency. In data centers, heat generated by GPUs is recovered by coolant in a cold plate and released into the outside air using chillers and cooling towers, so a lot of power is consumed not only by the GPU calculations but also by the cooling equipment.
In recent years, there has been growing interest in high-temperature water cooling using coolants at around 45°C. Using high-temperature coolants allows for efficient heat dissipation by leveraging the temperature difference with the ambient air, thus reducing cooling power consumption. However, with high-temperature water cooling, it is crucial that the coolant reaches the heat-generating components before it becomes excessively hot within the cold plate. Conventional cold plates typically have a structure where the coolant flows along the GPU surface, and because the coolant temperature rises towards the end of the flow path, it is difficult to uniformly supply low-temperature coolant to the heat-generating components.
Furthermore, in order to recover more heat from the GPU, it is necessary to divide the flow path into smaller sections to increase the heat transfer area. However, dividing the flow path into smaller sections makes it more difficult for the coolant to flow, and also increases pressure loss, which presents the challenge of requiring more power from the pumps that circulate the coolant.
Source: PRTimes
English