Skip to content
HolonomiX
Products/HX-AIFactoryTwin
Deployed

HX-AIFactoryTwin

Decision-grade thermal evidence for AI factories — before capital commits.

Runs, compares, validates, and attests 3D thermal-fluid scenarios for pods, halls, and datacenter layouts on NVIDIA GPUs. Signed evidence per run, verifiable offline.

StatusDeployed
ComputeNVIDIA A100 · H100
PhysicsNavier-Stokes · Boussinesq coupling
Production grid512³ – 1024³
ProcurementPrivate offer / pilot
Request → Compile → Execute → Validate → Certify → Evidence

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 production engine solves the incompressible Navier-Stokes equations with a Boussinesq buoyancy coupling on the GPU: first-order upwind advection, central-difference diffusion, and a direct spectral pressure-Poisson solve that returns a divergence-free velocity field. Racks enter as immersed boundaries with per-rack volumetric heat sources. Runs are deterministic.

How a run executes

A scenario compiles to an instruction program, executes on the GPU, passes validation, and leaves signed evidence behind.

Request →Contract resolve →Compile to IR →Execute on GPU →Physics QoI →Validation gates →Sanitizer →Certificate →Evidence bundle
Compiler

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.

Runtime

The production 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. Torch on CUDA, with a CPU fallback.

Honest scope

This is a thermal scenario evidence engine, not a wall-resolved CFD reference solver. It does not replace CFD.

Decision-QoI · Default

The thermal decisions teams actually need to make.

A fast production mode optimized for rack inlet temperature, hot aisle temperature, aisle ΔT, thermal excursion, recirculation, and heat removed against heat injected.

High-Fidelity Velocity

Stricter flow-field fidelity.

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.

Every run emits verifiable artifacts: configuration, GPU and container metadata, physics quantities of interest, validation status, checksums, and the signed certificate that binds them.

A run that fails validation cannot be attested

Four gates run on every scenario. They resolve to a tri-state verdict, and the verdict controls whether a certificate is issued at all.

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.

PASS Attestation proceeds.
WARN Recorded in the bundle and carried with the result.
FAIL Blocks attestation. No certificate is issued.

Fourteen files, checksum-addressed

A thermal claim that a third party can check offline, without calling HolonomiX.

manifest config input hashes result hashes validation certificate proof report QoI gate result raw telemetry checksums gpu info container info benchmark card

The bundle is Ed25519-signed, with an HMAC fallback. The verify-artifacts path recomputes every checksum, verifies the signature, and binds the result hash. Verification is fully offline.

No result overstates its fidelity

A native 512 cubed grid is roughly a 22 GB working set, which puts it on A100 and H100 class hardware. On a smaller device the runtime computes a memory-safe effective grid rather than failing or silently truncating.

Recorded per run grid_requested

What the scenario asked for.

Recorded per run grid_effective

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.

Benchmark card

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.

Preset hyperscaler_pod

40 racks, 0.46 MW

Backend staggered_mac / spectral_fft

Production backend

Grid 512³ → 64³

Requested to effective, memory-safe

Device NVIDIA RTX 5070 Laptop GPU

Dev class

Verdict pass

All four gates

Mean step time ~0.017 s/step

At 64³

Claim boundary Dev-class demonstration, not a production benchmark. Production numbers are measured on A100 and H100 at native 512 cubed and 1024 cubed, and are published as their own cards. 
hx-aifactorytwin run --preset hyperscaler_pod --steps 20 --evidence full --out ./out

Presets and deployment

Scenario presets run from single-row containment up to the H7-960 hall at 960 racks and 33 MW.

Container

A CUDA 12.8 base image, non-root, run with GPU passthrough.

Kubernetes

A Helm chart targeting the NVIDIA GPU Operator, with GPU resource limits, secret, persistent volume, and liveness and readiness probes.

Observability

A ServiceMonitor and Prometheus metrics endpoint.

Interfaces

FastAPI service and CLI, both against the same engine and the same evidence path.

Shipped, next, and research

Roadmap items are not production claims, and the page says which is which.

Now, shipped
  • GPU-native thermal-fluid engine on the staggered MAC backend with spectral pressure projection.
  • Presets from single-row containment to the 960-rack, 33 MW hall.
  • FastAPI service, CLI, Helm chart, CUDA container.
  • Signed evidence bundles and the release gauntlet.
Developer Edition 0.1.0.
Next
  • Native 512 cubed and 1024 cubed production proof set on A100 and H100, with published benchmark cards.
  • Container security scans and SBOM in continuous integration.
  • Multi-GPU job scheduling.
  • Campus compositor for multi-hall coupling.
0.2 to 0.3.
Research to roadmap
  • QTT research mode: experimental compressed-state projection, disabled by default behind an explicit flag.
  • QTT surrogate: compact physical-state serving and surrogate acceleration, targeting interactive, fleet-scale thermal twins.
The long-term differentiator and the basis of the acceleration thesis. Not a present-day performance claim, and not included in any measured number on this page.