ASRock Rack Launches 4U GPU Server with Eight NVIDIA Blackwell Ultra GPUs
Key Takeaways
- ▸The 8-GPU Blackwell configuration represents a proven topology for production AI workloads, balancing performance with established cluster management patterns
- ▸Compact 4U liquid-cooled design delivers significantly higher GPU density than air-cooled alternatives, critical for space-constrained data centers
- ▸6.4 Tbps of front-panel network bandwidth via OSFP ports enables high-speed GPU-to-GPU interconnect and node clustering without additional PCIe cards
Summary
ASRock Rack has unveiled the 4U16X-GNR2, a compact yet powerful GPU server featuring eight NVIDIA Blackwell Ultra GPUs, two Intel Xeon 6 'Granite Rapids' processors, and up to 12x 2.5" NVMe bays. The server is designed for high-performance AI training and inference workloads and includes advanced thermal management options, including liquid cooling and an optional two-phase ZutaCore cooling variant.
The system delivers 6.4 Tbps of network bandwidth on the front panel alone via eight 800Gbps OSFP ports, with an 8-GPU topology that represents a well-validated configuration for production AI deployments. The compact 4U form factor provides significantly higher density than traditional air-cooled systems, while redundant 3kW power supplies and hot-swappable components ensure reliability.
ASRock Rack offers flexible I/O and management configurations, with internal cabling options that allow customers to customize port placement without requiring separate SKUs. This modular approach reflects evolving data center infrastructure needs for large-scale AI deployment.
- Flexible I/O and thermal options allow customers to customize configurations for specific deployment scenarios without requiring different SKUs
Editorial Opinion
The ASRock Rack 4U16X-GNR2 exemplifies the engineering maturity now required for enterprise AI infrastructure. With eight Blackwell GPUs, advanced cooling, dense storage, and high-bandwidth networking in a compact chassis, this server reflects how hardware and software engineering must co-evolve to serve modern AI workloads. For data centers deploying large-scale AI models, such integrated systems are rapidly becoming the baseline—not premium—offering.
