ServeTheHome reported seeing ASRock Rack display a range of NVIDIA-based systems at OCP Global Summit 2026 in San Jose, from a 48RU Vera Rubin NVL72 rack to Rubin and RTX PRO GPU servers with different cooling designs. The clearest contrast is between the 2U and 5U HGX Rubin NVL8 configurations: they share an eight-GPU platform but differ in cooling, storage bays, expansion and power delivery.
Rack-scale compute and an eight-GPU Rubin system
The reported XGMRACK-VR/NVL72 combines 18 compute trays and nine NVLink 6 switch trays in a 48RU rack, with 36 Vera CPUs and 72 Rubin GPUs. ServeTheHome describes the trays as fully liquid-cooled and lists several rack cooling arrangements, including in-rack and in-row systems.
The same booth included the 2U16X-GNR2/DLC RUBIN, an HGX Rubin NVL8 system in a 2U chassis. Its reported liquid cooling covers the CPUs, GPUs and E1.S drives. The system has eight hot-swap E1.S NVMe bays and uses a 54V DC busbar for power.
How the 5U HGX Rubin differs
The 5U16X-GNR2/DLC RUBIN keeps the HGX Rubin NVL8 platform and dual Xeon 6 host approach, but uses hybrid cooling: liquid for the CPUs and GPUs, air for the rest of the system. Its larger chassis has 10 E1.S NVMe bays and four PCIe Gen6 x16 expansion slots. It uses conventional power supplies instead of the 2U system’s 54V DC busbar.
| System | Chassis | GPUs | Cooling | E1.S bays | Power architecture |
| 2U16X-GNR2/DLC RUBIN | 2U | 8 | Liquid cooling for drives, CPUs and GPUs | 8 | 54V DC busbar |
| 5U16X-GNR2/DLC RUBIN | 5U | 8 | Hybrid: liquid for CPUs and GPUs; air for the rest | 10 | Conventional power supplies |
In practical terms, the reported 2U layout concentrates on a fanless, fully liquid-cooled design, while the 5U version makes room for additional drive bays and PCIe expansion. The configurations make different tradeoffs in chassis space, cooling and power delivery.
Hybrid and air-cooled RTX PRO servers
ServeTheHome also described two MGX systems built around NVIDIA RTX PRO GPUs. The 6UXGM-GNR2 CX8/DLC supports up to eight RTX PRO 6000 Blackwell Server Edition GPUs. Its upper 2U is air-cooled, while the lower 4U is liquid-cooled.
The 4UXGM-GNR2 CX8 uses conventional air cooling and supports eight dual-slot PCIe Gen5 x16 GPUs. Together, the reported 6U and 4U designs show two more approaches in the booth lineup: a hybrid chassis for liquid-cooled GPUs and a separate air-cooled GPU server.
What the NVFP4 figures describe
ServeTheHome attributes the following NVFP4 platform figures to NVIDIA: 3,600 PFLOPS for sparse inference and 2,520 PFLOPS for dense training on the Vera Rubin NVL72, and 400 PFLOPS for sparse compute on HGX Rubin NVL8. These are reported platform figures, not benchmark results from measured runs.
