foxAIAIが動く。未来が更新される。
日本語
中文English日本語한국어

Liquid-Cooled NVIDIA Servers at the ASRock Rack OCP 2026 Booth

We checked out the ASRock Rack booth at OCP 2026 and saw servers ranging from fully integrated NVIDIA NVL72 racks to liquid-cooled switches The post Liquid-Cooled NVIDIA Servers at the ASRock Rack OCP 2026 Booth appeared first on ServeTheHome.

翻訳準備中のため原文を表示しています

Server Systems

At OCP Global Summit 2026 in San Jose, we stopped by the ASRock Rack booth and saw a broad set of NVIDIA systems. That included a full NVIDIA Vera Rubin NVL72 rack. Next to it were HGX Rubin NVL8 servers in both 2U and 5U, two NVIDIA MGX RTX PRO servers, and a dual NVIDIA Vera CPU server. We also spotted the liquid-cooled NVIDIA Spectrum-X SN6600-LD switch and the passively cooled NVIDIA RTX PRO 4500 Blackwell Server Edition GPU.

A big theme was liquid cooling. ASRock Rack used direct-to-chip liquid cooling on its rack-scale system and HGX servers, including an 8x RTX PRO 6000 PCIe server.

ASRock Rack XGMRACK-VR/NVL72 NVIDIA Vera Rubin NVL72

The XGMRACK-VR/NVL72 is ASRock Rack’s version of the NVIDIA Vera Rubin NVL72. It is a 48RU NVIDIA MGX rack with a 5000A busbar and a liquid manifold. Inside are 18 Vera Rubin compute trays and nine NVLink 6 switch trays, for a total of 36 NVIDIA Vera CPUs and 72 NVIDIA Rubin GPUs.

NVIDIA lists the full rack at 3,600 PFLOPS of NVFP4 inference (sparse) and 2,520 PFLOPS of NVFP4 training (dense). It has 20.7TB of HBM4, and each Rubin GPU gets 288GB. Power comes from four 3RU 19″ power shelves. Each shelf is rated at 110kW and uses six 18.3kW PSUs plus bulk capacitors. ASRock Rack offers in-rack, in-row, RDHx, or liquid-to-air sidecar CDU options.

Each compute tray is a 1U NVIDIA MGX design with two Vera Rubin Superchips. Each Superchip connects two Rubin GPUs to one Vera CPU over NVLink-C2C. Networking is eight 800Gbps OSFP ports from NVIDIA ConnectX-9 SuperNICs, plus two 400Gbps QSFP112 ports from an NVIDIA BlueField-4 DPU. There are even management switches up top for the lower-speed networking.

The trays are fully liquid-cooled, which helps lower the total power consumption in the rack, and is something that NVIDIA has been advancing along with its partners for a few generations now.

For those who want a more traditional HGX 8-GPU form factor, there were options for that as well, but with a twist.

ASRock Rack 2U16X-GNR2/DLC RUBIN with NVIDIA HGX Rubin NVL8

The 2U16X-GNR2/DLC RUBIN was one of the most impressive systems at the booth. It puts an NVIDIA HGX Rubin NVL8 board, with eight Rubin GPUs, into a 2U chassis. It is fully liquid-cooled, covering the E1.S drives, CPUs, and GPUs, with no fans. Networking includes NVIDIA ConnectX-9 SuperNICs and BlueField-4 DPUs. Power comes in over a 54V DC busbar.

The host is a dual-socket Intel Xeon 6 platform. It supports the Xeon 6700P, 6500P, and 6700E series in LGA 4710 sockets, with 16+16 DIMM slots at 2DPC. While it supports other series, given the cost of the rest of the system, we would imagine the vast majority of folks buying these will opt for the Xeon 6700P series.

There are eight 800Gbps OSFP ports via NVIDIA ConnectX-9, along with 8x hot-swap E1.S NVMe bays. NVIDIA quotes 400 PFLOPS of NVFP4 compute (sparse) for HGX Rubin NVL8.

The HGX 8-GPU platform underwent major changes in this generation, including a change in GPU orientation in the tray. This change enables it to work in high-density, liquid-cooled racks.

There was more in the booth, including hybrid-cooled servers, so let us get to those next.

情報源

ServeTheHome ↗