HGX H100 Explained: The Baseboard Behind Every 8-GPU Server
HGX H100 is the integrated GPU baseboard at the heart of every serious H100 server. Here is what it includes, how it differs from DGX, and why it matters when you buy.
When you buy an 8-GPU H100 server from any major vendor, the component doing the heavy lifting is the NVIDIA HGX H100 baseboard. It is the engineered foundation that turns eight individual SXM5 modules into a single coherent accelerator. Understanding what HGX is — and what it is not — helps you read server specs accurately and compare quotes that all claim to offer 'an 8x H100 system.'
What HGX H100 actually is
HGX H100 is a pre-integrated GPU baseboard that NVIDIA supplies to server makers. It carries the SXM5 H100 modules, the NVSwitch chips that wire them together, the NVLink fabric, and the power and signal delivery for the whole GPU complex. It is the GPU subsystem only — server vendors pair it with a host platform to make a complete machine.
- 4 or 8 H100 SXM5 modules (80GB HBM3 each), at up to 700W per GPU.
- NVSwitch chips providing a fully non-blocking 900 GB/s NVLink fabric between all GPUs.
- Integrated power delivery and the mechanical platform the host chassis is built around.
- Optional NVLink Switch System connectors for extending the fabric across nodes.
HGX is the GPU subsystem, not the whole server
The baseboard handles the GPUs and their interconnect, but a working server needs much more: host CPUs (typically dual Intel Xeon or AMD EPYC), system memory often in the 1-2TB range, NVMe storage, and high-speed networking — usually multiple ConnectX or BlueField adapters for InfiniBand or RoCE. Server vendors design that host platform around the HGX baseboard, which is why two '8x H100' servers can differ substantially in CPU, memory, storage, and network options even though the GPU subsystem is identical.
HGX vs DGX: integrated component vs full appliance
DGX H100 is NVIDIA's own fully-built, fully-supported appliance — a fixed configuration with its own software stack and support contract. HGX H100 is the baseboard that NVIDIA licenses to partners like Dell, HPE, Lenovo, and Supermicro, who build their own servers around it. The HGX path gives you choice in CPU, memory, storage, networking, and chassis, while DGX gives you a single standardized appliance. The GPUs and NVSwitch fabric are the same in both.
Why this matters when you buy
- An HGX-based server lets you tune the host platform — CPU core count, memory capacity, NVMe, and NIC choice — to your workload.
- Cooling design varies by vendor: some HGX chassis are air-cooled, others are built for direct liquid cooling.
- Compare the host platform, not just the GPU count, when you evaluate quotes for '8x H100' systems.
- Networking layout (number and speed of adapters) determines how well the node scales into a multi-node cluster.
Nexus Compute configures HGX H100 servers from the baseboard outward — selecting and validating the CPU, memory, storage, and networking that fit your workload — then tests and warranty-backs the complete system and returns a quote within 48 business hours.
Planning a hardware investment?
Tell us what you're trying to build. A procurement specialist will help you specify and quote the right configuration — within 48 business hours, no obligation.