The latest ASRock Rack 4U16X-GNR2 review from ServeTheHome puts numbers and port counts around a very specific kind of AI box: a 4U NVIDIA HGX B300 server built around two Intel Xeon 6 “Granite Rapids” processors and eight NVIDIA Blackwell Ultra GPUs.
ServeTheHome said ASRock Rack provided remote access after the site photographed the system in Taipei, Taiwan, and disclosed the review as sponsored. That matters because this is not a casual workstation with expensive cards shoved into risers. It is a dense, liquid-cooled server aimed at AI training and inference deployments where GPU interconnect, front-panel networking and serviceability are the point.
What is the ASRock Rack 4U16X-GNR2?
The ASRock Rack 4U16X-GNR2 is a 4U HGX B300 platform using eight NVIDIA Blackwell Ultra GPUs connected through NVIDIA’s onboard NVLink fabric, according to ServeTheHome. HGX systems use a baseboard-style GPU complex rather than ordinary add-in cards, which gives server makers a more standardized eight-GPU building block for high-bandwidth AI workloads.
ServeTheHome described the eight-GPU layout as a familiar topology for many AI jobs, which is the polite way of saying the software stack knows what to do with it. ASRock Rack offers two liquid-cooled versions of the machine: a direct liquid cooling model and a ZutaCore-cooled version using two-phase cooling for buyers that do not want water in the server loop.
Ports, storage and cooling
The front of the server carries 12 2.5-inch U.2 NVMe bays. ServeTheHome reported that two bays connect to the CPU side of the platform, while the other ten attach through the PCIe switch complex. That split is exactly the sort of detail that decides whether storage traffic fights with accelerators or stays where the system designer intended.
Front I/O includes four USB 3 Type-A ports, VGA and a power button. ASRock Rack also places the management port and two Intel i350-based 1GbE ports on the front, according to the review. If a deployment wants those links at the back, internal Ethernet cables can move connectivity to rear-facing ports, or split the arrangement between front and rear.
Next to the platform networking are the CPU liquid-cooling hose connections. ServeTheHome noted blue and red fittings for cooler incoming liquid and warmer outgoing liquid. The front panel also has eight OSFP cages rated at 800Gbps each, giving the server more than 6.4Tbps of network bandwidth before anyone adds a PCIe card. Yes, the face of the server is doing a lot of work here.
Why does the chassis design matter?
ServeTheHome called the 4U chassis compact compared with an air-cooled equivalent. That comparison is fair in this class: eight high-end GPUs, two server CPUs and high-speed networking generate enough heat that air cooling starts demanding more volume, more fan power and more tolerance for rack-level misery.
At the rear, the system includes ten 3kW 80Plus Titanium power supplies for redundancy, according to ServeTheHome. Cooling hardware includes large hot-swappable fan modules on each side and three smaller fans for the lower center section of the chassis. The review also showed rear management and USB access, plus the relocated 1GbE ports when the internal cabling is configured that way.
The useful read on this machine is less “ASRock Rack made a big GPU server” and more that the company is packaging HGX B300, liquid cooling and 800G-class networking into a serviceable 4U design. For AI operators, the interesting parts are the routing choices: which ports are up front, which SSDs hang off CPU lanes or PCIe switches, and how much cooling plumbing the rack team has to accept.
This story draws on original reporting from ServeTheHome.
