NVIDIA INCEPTION // THERMAL AI

Thermal intelligence
for data center cooling
at GPU scale.

We train transformer-based thermal forecasting and reinforcement-learning HVAC control models that reduce data center cooling energy by up to 40% — deployed from H100 training clusters to Jetson edge devices.

40%  cooling cost reduction PUE 1.15  target efficiency 200ms  control loop latency

Thermal models that learn
the physics of heat.

Our models combine physics-informed neural networks with sequence-to-sequence transformers. They learn the thermal dynamics of a data center — airflow, heat transfer, workload coupling — and predict temperature distributions across zones before they happen.

Model Architecture — ThermalNet v2.1
INPUT LAYER
Sensor Fusion Encoder
4,096 sensor channels · CRAC · IT load · ambient
EMBEDDING
Physics-Informed Transformer
12 layers · 8 heads · d_model=768 · Fourier positional encoding
FORECAST HEAD
Temporal Decoder
Horizon: 2–6h · 5-min intervals · 72 steps
34.2Mparams
INT8edge quant
14msinference

Architecture

ThermalNet processes live telemetry from thousands of IT, CRAC, and environmental sensors through a physics-informed transformer encoder. Fourier-domain positional encoding captures the periodic nature of thermal cycles. The decoder generates multi-zone temperature forecasts with calibrated uncertainty bounds.

Encoder12-layer Transformer, d=768, 8 heads
LossMSE + physical constraint regularization
UncertaintyMonte Carlo dropout, 50 samples
Training data10M+ hours sensor logs, 8 facilities
FrameworkPyTorch 2.4 · CUDA 12.5 · cuDNN 9

Three models. One thermal stack.

Each model is trained, validated, and deployed independently — then composed into a closed-loop control system. All models export to ONNX and run quantized on NVIDIA Jetson edge hardware.

Model Architecture Parameters Training Accuracy
ThermalNet
xkool-thermal-v2.1
Multi-zone thermal forecasting. Predicts temperature distributions across data center zones 2–6 hours ahead using a physics-informed transformer encoder with temporal decoder.
Transformer
12-layer encoder
8 heads, d=768
Fourier positional encoding
34.2M
F32 inference
INT8 edge quant
ONNX export
3,200 GPU-hrs
8× H100 SXM
batch=256
bf16 mixed precision
0.28°C
MAE @ 2h horizon
CoolRL
xkool-control-v1.4
Reinforcement-learning HVAC controller. Optimizes airflow, CRAC setpoints, and valve positions in real-time per zone. Reward: energy cost subject to thermal SLA constraints.
PPO + GNN
Actor-critic
Graph neural net
zone adjacency
61.8M
2 policies
action space=512
continuous control
8,400 GPU-hrs
8× H100 SXM
rollout parallelism
16 envs / GPU
35.4%
energy reduction vs baseline
WorkloadPredict
xkool-workload-v1.0
Forecasting model for workload-thermal coupling. Predicts GPU/CPU heat generation from job schedules, enabling pre-cooling of zones before high-compute jobs start.
Temporal Fusion
multi-horizon
attention + LSTM
covariate mixing
12.4M
F32 inference
INT8 edge quant
quantile output
1,100 GPU-hrs
4× H100 SXM
quantile loss
probabilistic
0.91
P90 quantile coverage

Training pipeline
from sensor to edge.

Models are trained on NVIDIA H100 clusters, validated against digital twins in Omniverse, and deployed as quantized INT8 graphs to Jetson Orin edge devices running inside the data center. The full loop — sensor read, inference, control action — completes in under 200ms.

01
Sensor Data Collection
Telemetry from IT load sensors, CRAC units, raised-floor pressure, and ambient monitors streamed at 1Hz across 4,000+ channels per facility. Aggregated into time-windowed training shards.
Kafka · Parquet · ~2.4 TB / facility / month
02
GPU Training
Distributed training on H100 SXM clusters. FSDP sharding across 8 GPUs, bf16 mixed precision, gradient checkpointing. CoolRL uses 16 parallel environment rollouts per GPU for PPO sample collection.
PyTorch 2.4 · CUDA 12.5 · NCCL · bf16
03
Digital Twin Validation
Trained policies validated against NVIDIA Omniverse digital twins of partner data centers. Simulates thermal response to control actions before any real-world deployment.
Omniverse · PhysX · CFD simulation
04
Edge Deployment
Quantized INT8 models deployed to NVIDIA Jetson Orin AGX devices on-site. TensorRT runtime, Triton inference server. Full control loop — sensor read, inference, HVAC actuation — in under 200ms.
TensorRT · Triton · Jetson Orin AGX 64GB

// Compute Requirements

Training cluster8× H100 SXM 80GB
Training hours (total)12,700 GPU-hrs
Per model retrain~3,200 GPU-hrs
RL rollout parallelism16 envs / GPU
Batch size (ThermalNet)256 · bf16
Memory per GPU72GB peak (FSDP)
Inference targetJetson Orin AGX
Control loop budget< 200ms end-to-end
InterconnectNVLink + InfiniBand
FrameworkPyTorch · CUDA 12.5
train_thermalnet.py
# ThermalNet v2.1 — distributed training on H100 cluster
import torch
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP

model = ThermalNet(
    d_model=768, n_heads=8, n_layers=12,
    sensor_channels=4096,
    physics_informed=True,
    uncertainty="mc_dropout",
)

model = FSDP(model, cpu_offload_policy=None)
optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4, weight_decay=0.01)

for epoch in range(100):
    for batch in dataloader:  # 10M+ hrs sensor data
        loss = mse_loss(model(batch.sensors), batch.temps)
        loss += physics_regularization(model, batch)  # heat eq. constraint
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

    if epoch % 10 == 0:
        export_onnx(model, quantize="int8")  # Jetson edge deploy

Benchmarked against
industry baselines.

Deployed across 3 partner data centers (totaling 42MW IT load). Results measured over 6-month controlled trials against standard PID-based cooling controllers. PUE measured per ASHRAE 90.4 methodology.

40%
Cooling energy reduction
avg across 3 facilities · 6-month trial
1.15
PUE achieved
vs 1.55 industry average (ASHRAE)
0.28°
Forecast accuracy (MAE)
2-hour horizon · zone-level prediction
Metric Baseline (PID) Xkool Delta
PUE (facility-wide) 1.55 1.15 −25.8%
Cooling energy (kWh/rack) 4,820 2,892 −40.0%
Thermal SLA violations 3.2 / week 0.4 / week −87.5%
Max inlet temperature 32.4°C 27.1°C −5.3°C
Forecast MAE (2h) 1.12°C 0.28°C −75.0%
Control loop latency 200ms real-time

Building the thermal
intelligence layer.

Xkool is an AI company focused on one problem: making data center cooling intelligent. We are a NVIDIA Inception member building the models, training infrastructure, and edge deployment stack for next-generation thermal management.

Data centers consume 1–1.5% of global electricity, and cooling accounts for up to 40% of that energy. As AI workloads scale exponentially, the thermal problem becomes a first-order constraint — not just on cost, but on compute density and reliability.

We are building the thermal intelligence layer: AI models that understand the physics of heat in a data center, predict thermal events before they occur, and control cooling infrastructure in real-time. Our models are trained on NVIDIA GPU clusters and deployed to edge devices running inside the facilities they optimize.

The result: 40% less cooling energy, PUE 1.15, and the ability to safely run higher-density racks without thermal risk — directly enabling the next generation of AI compute.

Founded
2024
Status
NVIDIA Inception Member
Headquarters
Remote-first
Focus
Thermal AI · Data center cooling optimization
Stack
PyTorch · CUDA · TensorRT · Triton
NVIDIA Omniverse · Jetson Orin