INTERACTIVE ARCHITECTURE STUDY

NVIDIA GB300 NVL72

Rack-scale compute. Clear fabric boundaries. One coherent view.

Plan your launch

Source-backed profilesVersioned designDistinct network planes

EAST–WEST / COMPUTE FABRICCanonical device groups · inspect for identities

Illustrative illumination · no live telemetry

Follow the work

Illustrative flow

Workload

Placement & failure study

Source domainDestination domainFailure

Place work deliberately.

DESIGN NOTE

Select equipment or a connection to inspect the complete identities. On a narrow screen, swipe the canvas or focus it and use arrow keys to pan.

Sources & assumptions ↗

Design for the planned maximum · growth & pods

Facility groups are site inputs, not vendor scalable units or network pods. Power, cooling, pathways and shared OOB dependencies still require site review.

Reference and assumptions · explicit gaps

Other accelerator references · TPU & AI ASIC

Vendor product references for an accelerator scope discussion. These need their own platform and interconnect models before an engineering study can be calculated.

Google Cloud TPU v6e / Trillium · cloud TPU service
AWS Trainium2 / EC2 Trn2 · cloud accelerator instances
Cerebras CS-3 / WSE-3 · wafer-scale system

Export this design · materials & wiring

Whole scoped model, fixed nine-sheet format. Draft gaps stay visible. Unknown cable assemblies, physical cages, lengths and prices remain unknown.

Design workbook · XLSX Tables & manifest · CSV ZIP

Platform configuration

System / fabric reference

GPUs required now Planned maximum GPUs

Accept displayed integral-equipment rounding

Loading the first-party reference catalog…

Switching tiers follow the selected system and fabric reference.

Selected study

Show

Logical relationships are not physical cage assignments, a purchasable cable schedule or measured performance.

DEVICE / CONNECTION INSPECTOR

Select any device

Select equipment or a wire bundle. Every aggregate accounts for its members.