# Datacenter 3D — source and fidelity ledger These are original, photo-informed visual reconstructions, **not manufacturer CAD, endorsed assets, or engineering drawings**. Three different platforms are represented: H100 SXM, HGX B300, and GB300 NVL72. The calculator also includes an illustrative eight-H100 server GPU tray; its separate fidelity and reference ledger is [H100-SOURCES.md](H100-SOURCES.md). The HGX B300 tray is NOT a compute tray inside the GB300 NVL72 rack, and the H100 is not a component of either Blackwell model. ## Rack — Supermicro SRS-GB300-NVL72 Primary source: https://www.supermicro.com/en/products/system/gpu/48u/srs-gb300-nvl72 Official product image inspected: https://www.supermicro.com/files_SYS/images/System/SRS-GB300-NVL72_main.jpg Cross-check: https://www.nvidia.com/en-us/data-center/gb300-nvl72/ Verified from Supermicro's product page: - 72 NVIDIA B300 GPUs and 36 Grace CPUs. - 18 × 1U ARS-121GL-NB3 compute nodes. - 9 NVLink switches. - 48U enclosure; width 600 mm × depth 1068 mm × height 2236 mm. - Eight power shelves and an in-rack liquid cooling distribution unit are described. Modeled: enclosure at the published dimensions, 18 compute faces arranged in two groups, nine central switch faces, eight six-module power shelves, a lower CDU, top networking and blank panels. The appearance and placement follow the official photo approximately. Inferred/simplified: precise rack-unit assignments, fascia geometry, fasteners, grille/port/drive detail and colors. Grilles are dark surface geometry, not physically drilled holes. Rear is a closed simplified panel. No internal GPU boards, fluid paths, wiring, fans, or thermal simulation. External handles/detail can project beyond the nominal enclosure bounds. Do not use total mesh bounds as installation-clearance measurements. **Power warning:** the source page lists “8x 1U 33kW (6x 5.5kW PSUs), total power 132kW,” which does not explain its redundancy arithmetic there. This is not evidence of measured operating consumption. Neither 132 kW nor the screenshot's illustrative 140 kW is used as a measured load in this pack. A calculator must keep operating-power assumptions separate and explicitly sourced. ## Tray — NVIDIA HGX B300, air-cooled reference appearance Visual identity source: https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/ The section explicitly titled **NVIDIA HGX B300** provides this official image: https://www.nvidia.com/content/nvidiaGDC/us/en_US/data-center/technologies/blackwell-architecture/_jcr_content/root/responsivegrid/nv_container_1760748008/nv_container/nv_teaser_copy_copy_.coreimg.jpeg/1766023108524/hgx-b300-nvl16-ari.jpeg Specification source: https://www.nvidia.com/en-us/data-center/hgx/ (Blackwell specification table, HGX B300 column). Verified: - Eight Blackwell Ultra SXM GPUs. - 2.1 TB total GPU memory, as published in NVIDIA's HGX B300 table. - Fifth-generation NVLink; 1.8 TB/s GPU-to-GPU bandwidth per GPU (not application throughput). Modeled from the labeled photo: four-by-two array of tall heatsink towers, warm metal top plates, dark side shrouds, dense horizontal fins, central finned brace, two lower front heatsinks, exposed baseboard and connector edge. Inferred/simplified: all tray dimensions, fin count, small connectors, PCB traces, component placement and materials. The approximately 0.49 × 0.76 × 0.23 m model is a visual scale assumption, **not a published product dimension**. GPU packages are obscured by heatsinks; they are not exposed dies or consumer graphics cards. The smaller heatsinks are not individually attributed to chips. No invented thermal or electrical connections are presented as facts. Top tick marks are model numbering, not vendor logos. The general HGX page's current hero image depicts a different, liquid-cooled layout. It was deliberately NOT used as the B300 visual reference. The historical-looking `nvl16` image filename is not used to infer GPU count; the page label and published HGX B300 table govern identity/count. ## Exposed module — NVIDIA H100 SXM Primary specification and visual source: https://www.nvidia.com/en-us/data-center/h100/ Official H100 images match the design supplied by the user: - https://www.nvidia.com/content/dam/en-zz/Solutions/gtcs22/data-center/h100/h100-og.jpg - https://www.nvidia.com/content/dam/en-zz/Solutions/gtcs22/data-center/h100/hopper-h100-kv-bb460_420-d.jpg Verified from the **H100 SXM** column, not the adjacent H100 NVL column: - SXM form factor. - 80 GB GPU memory. - 3.35 TB/s memory bandwidth. The page also lists configurable TDP up to 700 W. The combined calculator uses 0.7 kW as an editable H100-only scenario matching that maximum TDP, NOT typical measured consumption. The original asset-only preview did not use it as a calculator input. Modeled: a black module PCB, exposed central package, six visible surrounding package forms, banks of power components, paired copper mounting rings, metallic bridge components, small corner daughterboard, and a reflective gold/green/purple die motif. This is the bare-module **reference appearance**, not an operational system with cooling removed. No H100 activity lights or emitted die light are modeled. Inferred/simplified: the approximately 160 × 100 mm board scale, component identities and placement, ring dimensions, circuit traces, backside and silicon pattern. These are not manufacturer measurements, a transistor floorplan, or an electrical layout. Six visible package forms are reproduced from the photo; this does **not** assert six active HBM stacks. The die colors are artistic non-emissive metallic materials, not measured thin-film optics or temperature visualization. The mounting recesses are surface representations, not verified mechanical through-holes. Manufacturer imagery is not baked into textures. ## Rights and use Manufacturer photos were inspected as references but are **not bundled**, used as textures, or represented as licensed stock assets. All geometry and materials in the GLBs were built here. No manufacturer logos were reproduced. Product names identify the referenced products; trademarks remain their owners'. Citation is not a grant of manufacturer IP rights; review brand/product-image requirements before commercial publication. This is not a statement of legal clearance. ## Browser representation All four GLBs contain geometry and standard metallic/roughness materials, with no external textures, cameras, lights, compression decoders, or runtime scripts. The export is glTF 2.0, metres, Y-up. Named component nodes are preserved. Lighting and shadows are presentation effects, not physical cooling/power results. The original asset-pack preview adds steady power lights, slow pulses, and independent irregular rack activity blinks, without automatic rotation. Those animations and that preview's motion toggle are NOT included in the combined calculator: it uses static lights and manual-only rotation, with no automatic motion. Indicator lenses are modeled at 6 mm for readability, not from a verified lens dimension. They are illustrative, not measured telemetry or manufacturer status codes; no animation is baked into the GLBs. The calculator uses Google's `` 4.1.0 (Three.js underneath), vendored locally with its [Apache 2.0 license](vendor/model-viewer-LICENSE.txt). The bundle retains its embedded third-party notices. Math and labels are rendered independently in HTML; the local viewer and selected model load only when requested. Static transparent PNGs are included as fallback/poster images. ## Shipped files The four GLB/PNG pairs are copied unchanged from the approved prototype: `h100-sxm`, `h100-server`, `hgx-b300-tray`, and `gb300-nvl72-rack`. The prototype's Blender project, generator scripts, source photos and standalone HTML are not shipped. This ledger and the H100 ledger are adapted to describe this application, not a claim that publication rights have been independently reviewed. The route's `social.png` is a 1200 × 630 editorial cover composited from the approved `h100-server.png` poster with page typography in Chrome; it is not an additional manufacturer image or an unmodified Blender render.