AI large model training (e.g., GPT-5, ) and cloud inference computing centers are characterized by high-density cabinet deployment and continuous full-load operation.(e.g., NVIDIA B200 1800-2000W, and B300 >2500W). Traditional heatsinks fail to meet temperature control requirements. Skived fin heat sinks achieve ultimate cooling performance through the seamless integration of fins and substrate, perfectly solving the four core pain points of AI services: high heat flux, low thermal resistance, miniaturization, and high reliability. By fundamentally eliminating the contact thermal resistance of traditional heat sinks (Rth= 0.02°C/W), they break through the high-power dissipation bottleneck of AI. Having deeply penetrated the core links of the entire AI service industry chain, skived fin heat sinks have become the most mainstream application (GPU/CPU/TPU) in AI servers.

Integrated heat sinks employ monolithic forming processes, such as die casting, extrusion, and welding, to fabricate the cooling fins and substrate (or product housing) as a single, unified entity. This approach optimizes the thermal path from the design stage, minimizing the contact thermal resistance and space footprint inherent in traditional cooling solutions that rely on assembled components. Simultaneously, it enhances the overall structural integrity and reliability. Playing an irreplaceable role in applications like AI chips, servers, and computing clusters（NVIDIA H100/A100、AMD MI300）, integrated heat sinks have established themselves as one of the mainstream cooling solutions for high-end AI equipment.

The core challenge of AI servers lies in the surging power density of chips (single GPU power consumption exceeding 700W, heat flux density reaching 350~500W/cm²), high-density integration of computing power, and stringent requirements for long-term stable operation. Brazed cold plates, is tailored to meet the cooling demands of AI chips (GPU/TPU/NVIDIA H100/GB200) characterized by ultra-high power density and 7×24h uninterrupted operation.Brazed cold plates rapidly dissipate the heat directly from the chips through internal coolant circulation. This prevents excessive junction temperatures that could lead to computing power throttling or chip damage. As a key solution to overcome the bottlenecks of air cooling and unlock the full potential of AI computing power, brazed cold plates have become the mainstream standard for liquid cooling in high-end AI servers.

FSW water cooling, with its ultra-high sealing reliability and precise temperature control capability, ensures the temperature fluctuation of AI chips during full-load operation is controlled within an extremely small range, thus ensuring the stability and sustainability of the computing power output, significantly improving the training efficiency large models and the long-term service life of chips. It is the core supporting technology for AI service products to achieve "high density, low leakage, and long endurance" cooling By constructing an integrated sealed flow channel through solid-phase welding process and upgrading the detailed flow channel design (such as micro-channel array, partition temperature control flow channel, etc), its thermal exchange efficiency is more than 30% higher than traditional water cooling. It can efficiently cope with the dual challenges of ultra-high heat flow density and compact space of AI chips, high-density computing power clusters, and liquid cooling servers, and it is one of the key core components in the current all-liquid cooling computing center architecture.

The core requirements of AI servers include collaborative cooling of multiple chips, high-density computing power integration, and room-level large-scale deployment.Compared with traditional decentralized liquid cooling connection solutions, Manifold, which is tailored to the high-density, high-reliability and large-scale operation needs of AI servers, is integrated inside the chassis of AI servers. It connects multiple cooling terminals such as GPU/TPU cold plates, power supply cold plates and memory module cold plates. With precise flow distribution, it ensures consistent cooling performance across multiple chips. Its highly efficient and simplified connection and deployment scheme is well-suited for the high-density integration of AI computing power. Endowed with flexible and ultra-strong system scalability as well as reliable operation guarantees, Manifold addresses the thermal management challenges of multi-chip and multi-node systems, and serves as a critical component driving the large-scale and high-efficiency operation of AI computing clusters.