GB300液冷方案调研

Summary of Key Points from the Conference Call Industry and Company Involved - The discussion focuses on the liquid cooling solutions for the GB300 and GB200 systems, particularly involving companies like NVIDIA, Cooler Master, and AVC. Core Insights and Arguments 1. Changes in Liquid Cooling Solutions: The GB300 liquid cooling solution has significant changes compared to the GB200, particularly in the design of the cold plates and the use of quick connectors [1] 2. Quick Connector Configuration: Each compute tray in the GB200 system uses a liquid cooling design with a pair of quick connectors, with multiple loops connected through a manifold [2] 3. Quick Connector Pricing: The price of quick connectors in the GB200 system has decreased from approximately $100 to a range of $70 to $80 [6] 4. Supplier Determination: The quick connector suppliers are determined by the design party, with NVIDIA primarily using products from CPC and Staubli [7] 5. GB300 Design Changes: The GB300 features independent cold plates for each chip, requiring a total of 14 pairs of quick connectors per compute tray, leading to a total of 252 pairs for the NVL72 system [8] 6. New Quick Connector Pricing: The new NV UQD03 quick connectors in the GB300 have a reduced price of $40 to $50, significantly lower than previous models [8] 7. Initial Production Suppliers: Key suppliers in the initial production phase of the NV UQD03 include Cooler Master, AVC, and Readore, with Cooler Master taking a leading role [9] 8. Cooler Master’s Progress: Cooler Master has advanced quickly in the GB300 project due to its close collaboration with NVIDIA, which is responsible for the core design [10] 9. GB200 Project Participants: The GB200 project was initially led by Intel and several cloud service providers, with CPC being a significant player in the quick connector manufacturing [11] 10. Supply Chain Changes for GB300: The GB300 did not follow the GB200 supply chain due to the complexity of the new design and the need for more rigorous testing [12] 11. Current Supplier Progress: Cooler Master is leading the verification phase for the GB300 project, while AVC and Readore are still in the initial verification stages [13] 12. Cooler Master’s Supply Chain: The specifics of Cooler Master’s Tier 2 supply chain support remain unclear, with some factories located in mainland China [14] 13. UQD03 Design Rationale: The UQD03 was designed to be smaller to accommodate increased board slot numbers, but this may increase the risk of leakage due to tighter sealing requirements [15] 14. Significant Changes in GB300: The main changes in GB300 focus on the internal cold plate design, while other components like the manifold and CDU remain unchanged from GB200 [16] 15. Future Cooling Solutions: There is a potential shift towards water cooling for transceiver connectors, which could increase manufacturing complexity and costs [17] Other Important but Overlooked Content - The complexity of the assembly process has increased due to the smaller size of the quick connectors, which requires more precise assembly and additional reliability testing [12][15] - The transition to water cooling for transceivers is still in the design phase, indicating that the final implementation is not yet determined [17]