The GB300 is anticipated to be unveiled in the second quarter of 2025, with trial production starting in the third quarter. It will reportedly have a stronger cooling requirement, utilizing fewer motherboard fans, which implies an increased demand for liquid cooling solutions.
The supply chain revealed that the design of the GB300 has been underway for over two months. The motherboard design will shift from directly soldering the CPU and GPU onto the compute board to using slots for GPU installation, making it easier for customers to replace components. Additionally, the number of fans on the GB300 motherboard will be reduced.
One ODM noted that the number of fans used in the GB300 motherboard will be fewer than those in the GB200, primarily reducing the fans associated with the power supply and MOSFET sections, totaling a reduction of six fans. However, another ODM stated that the design of the GB300 motherboard is still subject to change, and the exact number of fan reductions remains undecided.
ODMs also mentioned that Nvidia is considering expanding the use of liquid cooling to memory modules, although this is still in the planning stage. There are two main reasons for this. First, the thermal management needs of memory do not yet warrant liquid cooling. Second, the development of liquid cooling for the GB200 has faced challenges. Nevertheless, from a server design perspective, reducing the number of fans seems to be the direction moving forward.
This does not mean that servers will no longer require fans. Currently, the shipment proportion of AI servers remains low, while non-AI servers predominantly rely on air cooling. Furthermore, even if the design of AI servers trends towards fewer fans, they will still primarily utilize water-to-air cooling systems in the next two to three years. The specifications for fans required by their Sidecar units will remain high in quantity, suggesting that the market value for fans is likely to increase.
The supply chain reported varying opinions regarding the thermal design power (TDP) of the B300 GPU. Some manufacturers claim that the TDP could reach 1,400W, significantly higher than the current range of 1,000W to 1,200W for the B200. Others assert that it will only rise to 1,200W, indicating a limited increase. Regardless of the version, there is consensus regarding the rising importance of liquid cooling solutions.
GB300 may require more complex liquid cooling routing
If the motherboard design of the GB300 changes, it will lead to more complex liquid cooling routing compared to the GB200. What was originally a single cooling loop for the compute board module will evolve into three loops to accommodate the slot-based design, complicating the layout and requiring more quick connect disconnect (QCD) connectors.
Cooling solution providers stated that various companies have begun implementing liquid cooling for the GB200, but most are still in a trial phase. Although the process has encountered significant issues including "flooding," Nvidia's trend toward dense designs for AI server racks suggests that liquid cooling will become essential.
Cooling module manufacturer Auras is optimistic about the liquid cooling market, estimating that revenue from liquid cooling will grow sixfold, with its share of total revenue projected to jump from 15% in 2024 to between 45% and 50% in 2025, effectively doubling overall company revenue. Supermicro CEO, Charles Liang, has also stated that liquid cooling is expected to achieve a 30% market share by 2025.