How has Supermicro Boosted Liquid-Cooled AI Data Centres?

Supermicro has announced expanded manufacturing capacity and advanced liquid-cooling technology to enable full-scale deployment of Nvidia Vera Rubin and Rubin AI platforms. 

Working closely with Nvidia, Supermicro delivers first-to-market rack-scale systems built for high-density, efficient data centre environments.

Using its modular Data Centre Building Block Solutions (DCBBS) and in-house cooling design, Supermicro offers operators scalable infrastructure to meet the increasing demand for AI and HPC.

Charles Liang, president and CEO of Supermicro, says: “Supermicro's long-standing partnership with Nvidia and our agile building block solutions enable us to bring the most advanced AI platforms to market faster than others.

“With expanded manufacturing and industry-leading liquid-cooling expertise, we're empowering hyperscalers and enterprises to deploy the Nvidia Vera Rubin and Rubin platforms infrastructure at scale with unmatched speed, efficiency and reliability.”

Liquid cooling for rack-scale GPU systems

Supermicro supports Nvidia’s flagship systems: the Vera Rubin NVL72 SuperCluster and the HGX Rubin NVL8 platform. These are designed for training and inference at massive scale.

The Vera Rubin NVL72 SuperCluster includes 72 Rubin GPUs, 36 Vera CPUs, Nvidia ConnectX-9 SuperNICs and BlueField-4 DPUs. 

All are connected by Nvidia NVLink 6, a high-speed link for GPU-to-GPU and CPU-to-GPU communication. It scales using Nvidia Quantum-X800 InfiniBand and Spectrum-X Ethernet.

This system delivers 3.6 exaflops of NVFP4 compute, 1.4 PB/s HBM4 bandwidth and 75 TB of memory. NVFP4 is a floating-point format for AI workloads, while HBM4 refers to high-bandwidth memory used in GPUs.

Supermicro builds the system into the third-generation Nvidia MGX rack architecture, designed for improved serviceability and uptime. Liquid cooling is handled by in-row Coolant Distribution Units (CDUs), which enable warm-water operation. This reduces energy use and water consumption while maintaining density and performance.

For enterprises with varied deployments, Supermicro offers a 2U liquid-cooled HGX Rubin NVL8 system. 

This compact rack unit features eight Rubin GPUs and delivers 400 petaflops NVFP4, 176 TB/s HBM4 bandwidth and 28.8 TB/s NVLink bandwidth. It uses 1600 Gb/s ConnectX-9 networking.

The system supports CPUs including next-generation Intel Xeon and AMD EPYC. Options include a high-density 2U busbar design for optimised rack integration, using Supermicro’s direct liquid cooling (DLC) technology to manage thermals at scale.

Platform features for AI model scale and protection

The Rubin platform includes a third-generation Transformer Engine, aimed at narrow-precision operations in long-context models. These are key for advanced AI training and large language models.

The Vera CPU uses Arm cores to double performance from the previous generation. It features 88 cores, 176 threads, 1.2 TB/s LPDDR5X memory bandwidth and 1.8 TB/s NVLink-C2C bandwidth to GPUs. Spatial multithreading improves concurrent execution of workloads.

Rack-scale confidential computing is enabled through Nvidia’s third-generation trusted execution environment. This isolates and protects models and data during processing.

Nvidia also includes a second-generation RAS (Reliability, Availability and Serviceability) Engine, performing real-time health checks to avoid system downtime and improve availability.

Networking comes from Nvidia Spectrum-X Ethernet Photonics, which introduces the Spectrum-6 ASIC. Built on a 3nm process with 200G SerDes and co-packaged optics, it supports 102.4 Tb/s switching with shared buffers. According to Nvidia this delivers "5x power efficiency, 10x reliability and 5x application uptime" over traditional optics.

Three switches are available: the SN6800 (409.6 Tb/s, 512x 800G ports, liquid-cooled), SN6810 (102.4 Tb/s, 128x 800G ports) and SN6600 (pluggable, 128x 800G ports in air or liquid-cooled models).

Supermicro also supports petascale all-flash storage and JBOF (Just a Bunch of Flash) systems with BlueField-4 DPU for a range of data centre storage workloads.

Manufacturing for AI-first data centre demand

Supermicro invests in expanded US-based manufacturing and a complete liquid-cooling stack to support Rubin and Vera Rubin platforms. Combined with its modular DCBBS architecture, this allows fast configuration, validation and scaling.

This setup gives hyperscalers and enterprises faster rollout and early access to high-density platforms optimised for modern AI infrastructure.