Simulate an AI data center
Model an AI data center's network, power, cooling and layout in software before hardware arrives, to find design conflicts early, test automation and plan how many GPUs fit within a fixed power budget.
The problem
AI data centers are expensive and slow to change once built. Rack power density, liquid cooling, cable routes and network topology all interact, and an error found during installation can cause long delays. Network automation and provisioning scripts usually cannot be tested until switches are racked.
Operators also need to know how many GPUs a site can run within its power envelope and how the facility should respond when the grid asks it to reduce load.
The approach
Simulation works at two levels. A logical simulation of the network reproduces switches, SuperNICs, DPUs and their software so teams can test topology, provisioning and automation; NVIDIA's DSX Air does this as a cloud-hosted service with a free trial. A physical digital twin models the building, racks, power and cooling; the Omniverse DSX Blueprint is NVIDIA's framework for AI factory twins.
DSX reference designs supply validated starting architectures, DSX MaxLPS covers power management within a fixed site budget, and DSX Flex handles grid signals. Access differs by part: reference designs need NVOnline access and NVIDIA offers Dynamic Power Software documentation for approved early-preview engagements, to scope site pilots.
For a few racks in an existing colocation hall, the provider's engineering review and standard design tools are usually enough. Full simulation pays off for new buildings, high-density liquid-cooled racks or large fabrics.1234
Conceptual architecture
Diagram as a list
Applications & solutions
- Facility digital twin (Omniverse DSX Blueprint)Models building, racks, power and cooling before constructionConnects to Power management model (DSX MaxLPS, Flex)
Operations & orchestration
- Provisioning and network automationScripts and tools tested against the simulated network
- Power management model (DSX MaxLPS, Flex)Plans GPU capacity within the power budget and grid limits
Accelerated computing
- Planned GPU racksCompute load that drives power and cooling demandConnects to Power management model (DSX MaxLPS, Flex), Planned fabric (Spectrum-X switches, SuperNICs, BlueField)
Networking, power & facilities
- Planned fabric (Spectrum-X switches, SuperNICs, BlueField)The network design under testConnects to Network simulation (DSX Air)
- Network simulation (DSX Air)Logical replica of switches, NICs, DPUs and their softwareConnects to Provisioning and network automation
Programs & resources
- Reference design (DSX)Validated starting architecture for compute, network, storage and facilityConnects to Facility digital twin (Omniverse DSX Blueprint)
Technologies and their roles
dsx4
Design, simulation and power platform
Reference designs, DSX Air, the Omniverse DSX Blueprint, MaxLPS and Flex cover the parts of a site that can be modeled before building.
omniverse5
Physical twin libraries
OpenUSD scenes, rendering and physics underpin the facility digital twin; several of these libraries are labeled pre-release and not enterprise-supported, so check each library's status and license.
spectrum-x3
Network being simulated
DSX Air simulates Spectrum-X switches, SuperNICs and BlueField DPUs with NetQ, so the fabric design can be tested first.
mission-control6
Operations after build
Applies job-aware power policies and health checks once the simulated design is running for real.
What you need first1
- Site drawings and power and cooling specifications
- Target rack design and network topology
- Network engineering and automation skills
- NVOnline access, arranged through your NVIDIA representative, for reference design documents
- 3D and OpenUSD skills for the facility twin, or a partner who provides them
Risks and how to reduce them
- Gaps between the model and the built site
- Use simulation as a design check, not a guarantee, and confirm with commissioning tests.
- Access-restricted or early-preview components
- Confirm which DSX parts you can use; reference designs need NVOnline access and Dynamic Power Software documentation is offered for approved early-preview engagements.
- Relying on vendor capacity scenarios7
- NVIDIA's power and capacity figures are scenario-based; run the numbers for your own site.
- The twin becomes outdated
- Update the model whenever the as-built site changes.
Related
Sources
- NVIDIA DSX documentation (opens in a new tab)
- NVIDIA Spectrum-X Ethernet Networking Platform (opens in a new tab)
- NVIDIA DSX Air Platform (opens in a new tab)
- NVIDIA DSX AI Factory Platform (opens in a new tab)
- NVIDIA Omniverse documentation (opens in a new tab)
- NVIDIA Mission Control (opens in a new tab)
- NVIDIA DSX MaxLPS overview (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.