As AI model training accelerates demand for computing power, traditional PCIe architectures can no longer meet high-performance bandwidth requirements. While NVIDIA’s SXM modules offer strong performance, adoption is limited by the proprietary NVLink protocol and high licensing costs. To address this, the Open Compute Project (OCP) introduced the OAM open standard, defining a universal AI accelerator interface that enables vendor interoperability and lowers costs. With advances in HBM3 memory and all-to-all interconnects, OAM is becoming the core form factor of next-generation AI servers. Against this backdrop, Gooxi has launched its flagship OAM server, “Haiqing,” built on open architecture to embrace diverse compute ecosystems.

Full-Stack Compatibility

Haiqing is designed on the OAM open standard, breaking through proprietary barriers and supporting a wide range of mainstream AI accelerators. It natively supports OAM 2.0 modules including: MetaX C550, Hygon Deep Compute 3, Iluvatar BI-V200, Kunlunxin P800, NVIDIA H200/B200 SXM, Intel® Gaudi® 3, and AMD Instinct™ MI325X. It also reserves compatibility for future domestic OAM modules, ensuring long-term investment protection and enabling enterprises to flexibly build unified training–inference heterogeneous computing pools with plug-and-play ease.

Hybrid Cooling for Extreme Power

With OAM modules now reaching 1,200W and approaching 1,400W, cooling has become a key bottleneck for AI cluster density. Haiqing adopts a 15×8080 high-speed front fan array with rear HIB board intelligent fan control, supporting up to 1,200W per module in air-cooled mode. It also integrates a liquid-cooling manifold, allowing seamless switching between air and liquid cooling. A tiered PID-based thermal control strategy delivers fast cooling under heavy loads, while prioritizing low noise and energy savings at lighter loads—optimizing PUE and advancing measurable green AI efficiency.

Lossless Interconnect

Haiqing integrates the motherboard and switch board with a high-density PCB design to enable lossless PCIe direct connections, eliminating cable-induced signal attenuation. Combined with PCIe lane aggregation and adaptive repair mechanisms, it ensures stable long-distance transmission. Dual front DPU support offloads networking, storage, and security tasks from the CPU, dedicating more resources to AI computation. Internal testing shows up to 3× performance gains over the previous generation in mainstream AI training, fully enabling training–inference integration.

Intelligent Operations & Maintenance

Haiqing extends reliability from hardware to software through a comprehensive intelligent O&M framework. Hardware features include dual-Flash operation and N+N redundant power. On the software side, it offers advanced diagnostics with one-click log collection, FMEA-based fault prediction, and SEL log visualization. Optimized Redfish and IPMI interfaces accelerate management response and support batch firmware upgrades and cluster-wide configuration, improving operational efficiency and stability across the data center.

Gooxi’s Haiqing server, built on the OAM standard, delivers lossless interconnects, hybrid cooling for extreme workloads, broad compatibility with domestic and international accelerators, and intelligent O&M for reliable cluster operations. By enabling seamless data flow and maximizing system efficiency, Haiqing achieves a leap from single-card performance to full system effectiveness.

The Haiqing server will be available for pilot testing later this year. Gooxi welcomes partners to schedule early access.