ASRock Rack's 4U16X-GNR2 server redefines "compact" AI infrastructure by integrating eight NVIDIA HGX B300 GPUs, two Intel Xeon 6 "Granite Rapids" processors, and liquid cooling into a 4U chassis. This dense configuration is designed for AI training, inference, and HPC clusters, positioning it as a scalable building block rather than an isolated appliance.

The core of the system is NVIDIA's HGX B300 platform, featuring a unified GPU fabric with up to 1.8 TB/s GPU-to-GPU bandwidth and 2.30TB of HBM3e memory per node. This high-bandwidth memory is crucial for large-language-model workloads that are often constrained by model size and context windows.

Fitting eight flagship GPUs and supporting components into 4U offers remarkable density, especially with liquid cooling. ASRock Rack offers two liquid-cooled variants: one with conventional direct liquid cooling (DLC) and another using Zuta Core's two-phase dielectric cooling. While liquid cooling addresses the primary heat load, substantial fan capacity remains essential for other components like memory, storage, and power supplies.