NVIDIA Vera HGX, Vera MGX, and NVIDIA RTX Pro 6000 Blackwell Server Edition Liquid-Cooled
Pegatron had both the HGX and MGX Vera platforms on display. HGX is designed for traditional NVL8 deployments, while MGX targets more modular configurations. The NVIDIA HGX Vera platform shown here is the foundation for 8-GPU AI servers. This baseboard has the Vera CPU, LPDDR5X memory, and then the high-density connectors to match the HGX platform’s OCP style-connectors.
The MGX Vera platform is a more modular approach that allows system integrators to build custom configurations around the Vera CPUs only, or also include Rubin GPU modules. This is a dual NVIDIA Vera module. Both CPUs get their LPDDR5X memory and then there are MCIO x8 connectors on the top and bottom of the board to provide PCIe connectivity to SSDs, NICs, GPUs, and other devices.
The NVIDIA RTX PRO 6000 Blackwell Server Edition with liquid cooling is designed for dense multi-GPU deployments. Currently this is known as the “BSE” version of the card. This is new for the show and integrates liquid cooling, and bus-bar power (instead of cables) to the popular card.
On the subject of the PCIe GPUs, let us get to another server with them.
Pegatron AS205-2T1
The AS205-2T1 is a 2U server that balances density with flexibility. This system supports NVIDIA RTX PRO 6000 GPUs and is designed for AI inference and graphics workloads. Here is the placard:
The front view of the AS205-2T1 shows a compact 2U design. The drive bay configuration adopts E1.S SSDs for compactness over capacity.
Inside, we get dual Intel Xeon 6 CPUs. This is an NVIDIA MGX/ OVX style design powered by Xeon.
We saw this view when we were looking at the 2.4kW PSUs so it appears as though Chenbro makes this chassis.
Next, let us take a quick look at a Rubin NVL8 platform.
Pegatron AS210-2T1-8H3
The AS210-2T1-8H3 is a 2U AI GPU server designed for high-density deployments. This system supports 8 GPUs in a compact chassis, making it suitable for edge AI and inference workloads.
The front view shows the dense configuration with drive bays and cooling intake. Fitting 8 GPUs into a 2U chassis requires careful thermal design. That includes using E1.S storage bays to keep the designs on 1U trays.
There is a spot for the NVIDIA BlueField-4 DPU as well.
Even with all of the advancements, we still get fairly traditional local I/O.
The idea here is that the 8x NVIDIA HGX Rubin GPUs get both NVLink via onboard NVLink switches and also get up to 2.3TB of HBM to run large models. Rubin’s HGX NVL8 board has integrated ConnectX-9 800Gbps ports with one for each GPU.
Pegatron also had a number of models with E3.S EDSFF storage, so let us look at those next.
