Pegatron's GTC 2026 booth showcased the liquid-cooled NVIDIA Vera Rubin NVL72 rack system with 72 GPUs and 36 CPUs, alongside PCIe GPU servers and networking innovations.
At NVIDIA GTC 2026, the Pegatron booth was packed with cutting-edge server solutions, from the massive NVIDIA Vera Rubin NVL72 rack to compact 2U Rubin NVL8 systems and high-density PCIe GPU options. The star of the show was undoubtedly the liquid-cooled SVR rack solution built around the NVIDIA Vera Rubin NVL72 architecture.
This rack-scale system unifies 72 Rubin GPUs with 36 Vera CPUs into a single platform designed for AI factory workloads. The design integrates high-speed NVLink switches through the center and rear spine, with power shelves positioned at the top and bottom, plus ancillary management networking. Unlike typical trade show displays where these systems appear as complete racks, Pegatron gave attendees a closer look at the internal components.
The RA4803-72N3 compute node represents Pegatron's implementation of the NVIDIA Rubin NVL72 compute tray. This node features optimized liquid cooling paths specifically engineered for the high-density GPU configuration. The front faceplate shows a new layout prioritizing both liquid cooling efficiency and serviceability, maintaining the dense compute configuration that makes NVL72 so powerful.
The front panel houses the NVIDIA ConnectX-9 SuperNIC, providing the networking backbone for Rubin systems. This component delivers up to 800Gb/s of bandwidth per port - critical for keeping all 72 GPUs in an NVL72 rack fed with data. The SuperNIC also includes an E1.S SSD and features a special connector at the rear for easy assembly and servicing. The entire design is optimized for liquid cooling, reflecting the thermal challenges of such dense configurations.
In the center of the node, additional networking cages house the NVIDIA BlueField-4 DPU. This infrastructure processor handles networking, storage, and security workloads, offloading these tasks from the main CPUs. The BlueField-4 is becoming a standard component in AI infrastructure to maximize compute efficiency. NVIDIA is particularly leveraging this DPU to accelerate KV caches in future generations, which explains why the CPU in this generation is significantly larger than previous iterations.
Beyond the NVL72 rack, Pegatron also displayed PCIe GPU servers, though specific details weren't provided in the booth coverage. These systems likely cater to different market segments requiring GPU acceleration without the extreme density of NVL72 configurations.
The Pegatron showcase at GTC 2026 demonstrates the rapid evolution of AI infrastructure, where liquid cooling, 800Gbps networking, and specialized infrastructure processors are becoming essential components rather than optional upgrades. The Vera Rubin NVL72 represents a significant leap in AI factory design, packing unprecedented compute density into a single rack while maintaining serviceability and thermal management.
Comments
Please log in or register to join the discussion