# H100 eight-GPU server tray — sources and fidelity ledger

This asset is an original, photo-informed visual reconstruction made for the approved GPU calculator prototype and reused in the GPU power cost calculator. It is **not manufacturer CAD, an engineering drawing, a thermal model, an endorsed asset, or a dimensional reference**.

## Primary NVIDIA references

- [NVIDIA HGX AI Factory components](https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory-h100-h200-b200/latest/components.html) identifies the HGX H100/H200 baseboard as an eight-GPU platform and shows an official HGX H100/H200 baseboard photograph.
- [Official HGX H100/H200 baseboard image](https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory-h100-h200-b200/latest/_images/hgx-ai-factory-h100-h200-b200-01.png) was inspected for the older-generation, low-profile, silver finned-heatsink visual language. The image is not bundled and is not used as a texture.
- [NVIDIA technical introduction to HGX H100](https://developer.nvidia.com/blog/introducing-nvidia-hgx-h100-an-accelerated-server-platform-for-ai-and-high-performance-computing/) states that the HGX H100 8-GPU platform hosts eight H100 Tensor Core GPUs and four third-generation NVSwitches.
- [NVIDIA H100 product page](https://www.nvidia.com/en-us/data-center/h100/) is the primary H100 generation and product-identity reference.

## OEM cross-check

- [Supermicro NVIDIA-Certified Systems](https://www.supermicro.com/en/accelerators/nvidia-certified-systems) lists the SYS-821GE-TNHR with an NVIDIA SXM HGX H100 8-GPU configuration. This confirms that an eight-H100 HGX baseboard is integrated into OEM server systems; the modeled tray is not a reconstruction of that server chassis.

## What is verified

- Product family: NVIDIA H100 / Hopper generation.
- Eight H100 GPUs on the referenced HGX H100 8-GPU platform.
- Four third-generation NVSwitches at the generation-level platform description.
- Official reference imagery uses low-profile metallic finned module covers that are visibly different from the taller Blackwell/B300 heatsink towers in the existing asset pack.

## What is modeled

- Eight individually named GPU-module mesh nodes in two rows of four.
- Short cool-silver fins with blue-gray top caps, dark insets, mounting hardware, and small abstract label marks.
- A blue-black baseboard, a shallow charcoal tray, four low switch-region heatsinks, structural rails, edge connectors, and representative small components.
- A transparent PNG poster rendered from an isolated orthographic studio camera.

## What is inferred or simplified

- The exact two-row-by-four placement is the requested comparison layout. It is not asserted to reproduce every HGX H100 production baseboard routing or mechanical revision.
- All dimensions, fin count, heatsink profiles, rail geometry, connector placement, fasteners, colors, PCB traces, and small component identities are inferred for presentation.
- Switch heatsinks identify a simplified region corresponding to the published four-NVSwitch platform fact; they are not a verified package or cooling layout.
- No fluid paths, airflow, cabling, back side, electrical connectivity, thermal behavior, or service clearances are modeled.
- No manufacturer logos, source photographs, or image textures are included. Abstract top marks provide visual orientation only.

## Scale and capacity warning

The model uses an approximate visual envelope of about 0.50 × 0.74 × 0.11 m so it sits comfortably beside the existing reconstructed tray assets. **That envelope is not a published HGX H100 dimension. Physical size, volume, heatsink height, or visual mass does not encode or imply compute capacity, memory, bandwidth, power draw, cooling requirement, throughput, or installation clearance.** Keep any calculator values independent, explicit, and separately sourced.

## Rights and intended use

Manufacturer pages and images were inspected as references only. All geometry and materials in the deliverable were authored procedurally for this prototype. Product names identify the referenced hardware; trademarks remain with their owners. Citation is not a grant of manufacturer IP rights, and this ledger is not a statement of legal clearance for publication.

## Suggested viewer framing

- `camera-orbit="-35deg 61deg 92%"`
- `camera-target="0m 0.055m 0m"`

The GLB is glTF 2.0 with an embedded binary buffer, Y-up conversion on export, standard metallic/roughness materials, named component nodes, and no textures, cameras, lights, compression decoders, or runtime scripts.