Available now · Private Appliance · Scoped Pilot
HX-AIFactoryTwin
Decision-grade thermal evidence for AI factories before capital commits.
Run and compare 3D thermal-fluid scenarios for pods, halls, and datacenter layouts on NVIDIA GPUs. Validation gates control whether a run is attested. Signed evidence can be verified offline.
Evaluate HX-AIFactoryTwin for your team
- Who it serves
- AI infrastructure and thermal engineering teams planning rack layout and cooling.
- When to evaluate
- You need to compare thermal scenarios before committing to a cooling or layout decision for an AI-factory pod, hall or datacenter.
- Outcome to establish
- Thermal quantities and run records documenting the actual execution profile, model assumptions and validation outcome, with signatures for eligible attestations.
- Deployment
- Private Appliance · Scoped pilot
- Evidence to review
- Signed execution evidence binds runtime profile, release, inputs and recorded outcome; Developer Edition benchmark cards identify requested and effective grids.
- Commercial starting point
- Scope a Private Appliance or pilot evaluation around representative geometry, operating conditions and reference data. Agree the qualified hardware profile, tolerances and treatment of warnings or guard stops.
Evidence scope: the public 64³ card describes the Developer Edition reference path. The qualified QTT-native runtime has its own H100 80 GB profile and qualification records. Review the runtime, scenario and evidence together.
Qualified QTT-native execution on H100
HX-AIFactoryTwin's qualified single-GPU H100 runtime supports QTT-native heat-MPO execution through a dense-equivalent 4096³ problem scale under the documented H100-80GB production profile. The runtime uses adaptive grouped and streaming hybrid truncation, emits offline-verifiable execution evidence, supports checkpoint-resumable operation, and fails closed when the qualified resource or runtime envelope is not satisfied.
Production profile
Execution is bound to a single exclusive H100 80 GB GPU with MIG disabled, the pinned runtime environment and the selected signed profile. The qualified profiles cover 512³, 1024³, 2048³ and 4096³ dense-equivalent scale. Admission rejects incompatible resources before allocation.
Qualification basis
The July 2026 qualification compares streaming-hybrid execution with an exact-gesvd reference from the same checkpoint. It establishes numerical equivalence within the stated float64 envelope, preservation of the failure boundary and memory-safe execution. It does not establish a speedup at 4096³ or empirical accuracy for every thermal scenario.
Evidence access
Request the release 0.1.0 qualification material, the applicable H100 profile and matching verification instructions for an evaluation. The reference and hybrid bundle identifiers are 600fb0f9a8c01564 and 4cb6643f3c88f115. These qualification records are distinct from the public development card below.
Production execution evidence binds the profile, release, input digests, checkpoint chain, backend, resource accounting and recorded result type. It uses Ed25519 certificates, with a labeled shared-secret HMAC fallback. Confirm the signing mode and expected issuer material supplied for the evaluation. The H100 production profile excludes multi-GPU execution, automatic scale-out and silent device or algorithm substitution. A physics-guard stop remains a recorded stop, not a successfully completed simulation.
Evaluate thermal risk, cooling strategy, layout changes, and failure scenarios before committing capital to physical infrastructure.
HX-AIFactoryTwin runs, compares, validates, and attests 3D thermal-fluid scenarios for AI-factory pods, halls, and datacenter layouts on NVIDIA GPUs. Infrastructure, facilities, and engineering teams evaluate rack heat, hot and cold aisle behavior, cooling effectiveness, obstruction effects, recirculation risk, thermal excursions, and heat-removal balance, per pod, per hall, per scenario.
The Developer Edition reference solver uses incompressible Navier-Stokes with Boussinesq buoyancy coupling: first-order upwind advection, central-difference diffusion and a direct spectral pressure-Poisson solve. Racks enter as immersed boundaries with per-rack volumetric heat sources. The qualified H100 runtime uses QTT-native heat-MPO execution under the separate profile described above; the two paths have distinct benchmark and validation records.
How the reference solver executes
The Developer Edition scenario path compiles an instruction program, executes on the GPU and records validation results. Certificate issuance depends on the validation verdict.
The datacenter compiler emits a 32-register instruction program: Boussinesq Navier-Stokes in three dimensions plus energy, an IMEX-CNAB2 time structure, Chorin projection on the staggered MAC backend, immersed-boundary rack masks, and separable initialization. Grid discipline holds a 512 cubed minimum.
The reference backend integrates incompressible Navier-Stokes with Boussinesq coupling: monotone first-order upwind advection, central-difference diffusion, a direct spectral pressure-Poisson solve giving Leray projection, a smooth immersed rack mask, per-rack volumetric heat sources, and two-thirds-rule dealiasing. It runs with Torch on CUDA. The qualified H100 production profile does not permit silent device substitution.
This is a thermal scenario evidence engine, not a wall-resolved CFD reference solver. It does not replace CFD.
Thermal decision mode
A scenario mode focused on rack inlet temperature, hot aisle temperature, aisle ΔT, thermal excursion, recirculation, and heat removed against heat injected.
High-Fidelity Velocity mode
For diagnostic, audit, and research use, where the flow field itself is the object of study.
The product ships as a containerized CUDA runtime with a FastAPI service, a CLI, a Helm chart, scenario presets, validation gates, and benchmark reporting.
The Developer Edition reference path retains validation artifacts and issues an attestation only for eligible results. The qualified H100 runtime signs execution evidence with its recorded outcome, including a guard stop. A signature does not convert a stopped run into a completed simulation.
A reference run that fails validation cannot be attested
The Developer Edition reference path applies these four gates. They resolve to a tri-state verdict that controls certificate eligibility. The qualified H100 runtime adds its profile-specific resource admission and physics guards; its execution evidence must preserve a guard stop as a stop.
Finite field
Every field value is finite. No silent NaN propagation.
Temperature bounds
The temperature field stays within -20 °C to 100 °C.
Divergence
Dimensionless divergence sits below the gate, so the velocity field is incompressible to tolerance.
Poisson residual
The spectral pressure-Poisson residual sits below tolerance.
The reference evidence bundle contains fourteen files
An attested Developer Edition run packages the following artifacts for offline review with the matching verifier and expected issuer key.
The reference bundle supports Ed25519 signatures and a shared-secret HMAC mode. Independent public-key verification uses Ed25519; HMAC verification requires the shared secret. The verify-artifacts path recomputes every checksum, verifies the signature, and binds the result hash. Verification is fully offline.
No result overstates its fidelity
The Developer Edition reference path records the requested and effective grids, including an explicitly reduced grid on smaller development devices. Its historical native 512³ A100 run has its own evidence. The qualified QTT-native H100 profile enforces its resource envelope before allocation rather than silently changing the grid or device.
What the scenario asked for.
What the device actually ran.
Both values travel in the evidence bundle. A reader can always tell a dev-class demonstration from a production run.
Public development card and native execution records
The card format exists so a dev-GPU demonstration can never be read as a production benchmark.
The authoritative machine card is produced by an actual run, and it records wall time, mean step time, validation verdict, backend, and the requested against effective grid.
40 racks, 0.46 MW
Developer Edition reference backend
Requested to effective, memory-safe
Dev class
All four gates
At 64³
hx-aifactorytwin run --preset hyperscaler_pod --steps 20 --evidence full --out ./out
Presets and deployment
Developer Edition reference presets run from single-row containment up to the H7-960 hall at 960 racks and 33 MW. Production evaluations bind scenario inputs to the selected qualified profile and release.
The Developer Edition reference image uses CUDA 12.8 with GPU passthrough and a non-root process. The qualified H100 profile pins its own image and CUDA 12.9 environment.
A Helm chart targeting the NVIDIA GPU Operator, with GPU resource limits, secret, persistent volume, and liveness and readiness probes.
A ServiceMonitor and Prometheus metrics endpoint.
FastAPI service and CLI, both against the same engine and the same evidence path.
Release scope and further development
Release 0.1.0 has a qualified QTT-native H100 runtime and retained Developer Edition reference measurements. Evidence must identify which path ran.
Qualified production
QTT-native heat-MPO execution on the documented exclusive H100 80 GB profile, adaptive grouped and streaming hybrid truncation, checkpoint-resumable operation, offline-verifiable execution evidence and release-aware rollback.
The signed profile and release determine supported hardware, software and resource limits.
Developer Edition reference
The staggered-MAC solver, spectral pressure projection, scenario presets, FastAPI service, CLI, Helm chart, CUDA container and reference validation suite remain distinct reference material. The public 64³ card and recorded native 512³ A100 runs belong to this path.
Further development
Fleet-scale surrogate serving and campus composition remain separate development directions. Multi-GPU execution and automatic scale-out are outside the current H100 production profile. Other devices and runtime combinations require separate qualification.
The qualified QTT-native runtime is a completed capability; it is distinct from future surrogate acceleration.
Preparing a thermal evaluation
Select a decision to evaluate, a representative scenario and the measurements needed to judge its result. Request the matching scenario contract, release profile and evidence material before preparing a full model.
Scenario inputs
Specify hall dimensions, rack positions and heat loads, cooling-unit locations and supply conditions, containment assumptions and the scenario changes to compare. Identify the reference data and quantities of interest, such as rack inlet temperature or heat-removal balance.
The Developer Edition includes a four-rack hot-aisle example and an H7-960 contract pack with a JSON schema, base geometry, scenario overrides and request examples.
Environment and integration
For qualified execution, reserve an exclusive H100 80 GB GPU and agree the signed runtime profile, container image and checkpoint/evidence storage. Confirm the API or CLI interface and any geometry conversion against the release supplied for the evaluation.
Reference preset timings do not establish that a customer scenario has been validated.
Acceptance criteria
Agree expected thermal quantities, reference comparisons, tolerances and the treatment of warnings or guard stops. Inspect the actual backend, grid, resource profile and validation outcome in the evidence. Verify the resulting bundle with the matching release tools and expected issuer material.
Thermal model requirements
Cooling and layout decisions for AI-factory pods, halls and datacenters, where scenario geometry, operating conditions and model fidelity must be assessed before committing capital.
Inputs
Scenario geometry, boundary conditions, quantities of interest, and available reference data.
Outputs
Thermal quantities of interest and run records documenting model assumptions, validation results, and the effective grid.
Limits
The public 64³ card documents the Developer Edition reference path. Qualified QTT-native execution is bound to the documented exclusive H100 80 GB profile; model accuracy and acceptance require scenario-specific evidence.
Availability: Private Appliance · Scoped pilot.