Skip to explorer
HOW COMPUTE WORKSAll experiences

4U LIQUID-COOLED REFERENCE SERVER

Supermicro HGX H100 server

A 4U liquid-cooled server with eight NVIDIA H100 GPUs on one HGX board: the building block of Colossus’s racks.

More about this server

Supermicro’s 4U liquid-cooled GPU server holds an NVIDIA HGX board with eight H100 GPUs linked by NVLink and NVSwitch, plus two Intel Xeon host CPUs and 32 memory slots. Cold plates on the chips carry heat into a liquid loop.

When

The Supermicro and ServeTheHome case study of December 2024 describes xAI using this 4U Universal GPU system in Colossus. The model here follows Supermicro’s reference datasheet and manual; the exact deployed SKU is not published.

Why it matters

Eight of these servers and a coolant distribution unit make one Colossus rack of 64 GPUs. The server is where GPUs meet the CPUs, network adapters and liquid cooling they depend on, the role the GB300 compute tray plays in newer NVIDIA racks.

Key facts

GPUs
8 NVIDIA H100 SXM (HGX)
CPUs
2 Intel Xeon Scalable
Cooling
Direct-to-chip liquid cold plates
Form factor
4U (SYS-421GE-TNHR2-LCC reference)

Supermicro · SYS-421GE-TNHR2-LCC datasheetSupermicro / ServeTheHome · xAI Colossus case study, December 2024Sources reviewed October 28, 2026

TEACHING MODELDrag to rotate

Opening the 3D model…

ILLUSTRATIVE MODEL · DRAG TO ROTATE
Follow heat or data

CONNECT THE SCALES

What moves through the system?

Sources & model notes · reviewed October 28, 2026

Separate host and HGX trays follow the reference service manual; not a verified deployed SKU or exact board layout. Exploded spacing is illustrative.

Supermicro · Reference server service manual ↗

Revision 1.0a; retrieved October 28, 2026 · Figures 1-1, 1-3, 1-5 and 1-6 (pages 10, 13, 15–16)

Reference front/rear panels and separate host/HGX trays. This does not establish the deployed Colossus SKU.

Supermicro · SYS-421GE-TNHR2-LCC reference ↗

Retrieved October 28, 2026 · Page 1, enclosure and interconnect

449 × 174 × 895.35 mm reference enclosure. Exact deployed SKU and internal placement are not established; H100 configuration selected.

NVIDIA · Hopper architecture ↗

March 22, 2022 · Figure 1 and H100 SXM5 configuration

One GH100 compute die, five enabled HBM3 stacks; schematic layers and connection paths.

Original authored geometry. No private CAD, exact facility floor plan, performance simulation, or current deployment total is implied. Cooling and networking lines show relationships, not surveyed routes.