Use Cases — CoolX Hyperscale Cooling

AI Training

GPU Cluster Density at Scale

More GPUs per rack. Zero thermal compromise.

Large-scale AI model training is the most thermally demanding workload in modern computing. Dense GPU clusters running sustained matrix operations push processor packages to their thermal limits continuously — and legacy air cooling forces operators to under-provision rack density to maintain headroom. CoolX eliminates that tradeoff entirely.

The Challenge

A leading AI research lab was running H100 GPU clusters at 60% of theoretical rack density due to thermal constraints. Peak training runs triggered throttling events that extended job completion times by 18–22% and introduced non-determinism in gradient accumulation.

The Outcome

After CoolX deployment, the lab achieved 3× rack density with zero throttle events across a 90-day continuous training campaign. Job completion times dropped by 19% on average. The thermal model predicted and neutralized 847 potential heat spikes during the period.

Results

3×GPU density per rack

19%Faster job completion

847Heat spikes neutralized

0Throttle events

Thermal Event Timeline

Legacy: throttle detected

Legacy: recovery time

CoolX: forecast horizon

CoolX: event neutralized

Relative time → CoolX acts before legacy systems detect the event

Hyperscale Inference

Inference at Hyperscale

Consistent latency. No thermal variance.

Inference workloads are latency-sensitive in a way training is not. A single throttle event on a serving cluster introduces tail latency spikes that cascade through dependent services — degrading user experience and violating SLAs. CoolX's predictive model is particularly effective here because inference workloads have highly regular thermal signatures that the AI can forecast with exceptional accuracy.

The Challenge

A Tier-1 cloud provider's inference fleet was experiencing p99 latency spikes of 40–80ms during peak traffic periods. Root cause analysis traced 73% of spikes to CPU thermal throttling on the serving nodes — a problem their existing cooling infrastructure couldn't address without a full rack redesign.

The Outcome

CoolX drop-in installation across 200 racks eliminated thermal-induced latency spikes entirely. p99 latency improved by 34ms. The provider reported a 28% reduction in over-provisioning costs as they could now run serving nodes at full utilization without thermal risk.

Results

34msp99 latency improvement

28%Less over-provisioning

200Racks deployed

73%Spike root cause eliminated

Thermal Event Timeline

Legacy: throttle detected

Legacy: recovery time

CoolX: forecast horizon

CoolX: event neutralized

Relative time → CoolX acts before legacy systems detect the event

PUE & Efficiency

PUE Reduction at Hyperscale

Targeted cooling. Dramatic efficiency gains.

Data center energy efficiency is measured by Power Usage Effectiveness — the ratio of total facility power to IT equipment power. Traditional cooling systems overcool entire rows or rooms to guarantee thermal safety margins, wasting enormous amounts of energy. CoolX's per-chip targeting eliminates overcooling waste at the source.

The Challenge

A colocation operator was running a PUE of 1.52 across their primary facility — above industry average and increasingly costly as energy prices rose. Their cooling infrastructure was designed for worst-case thermal loads and ran at full capacity regardless of actual workload.

The Outcome

CoolX deployment across the facility's highest-density zones reduced overall PUE to 1.09 within 90 days. Annual energy cost savings exceeded $4.2M. The operator has since committed to CoolX as the standard for all new rack deployments.

Results

1.09PUE achieved

40%Cooling energy reduction

90Days to full benefit

$4.2MAnnual savings

Thermal Event Timeline

Legacy: throttle detected

Legacy: recovery time

CoolX: forecast horizon

CoolX: event neutralized

Relative time → CoolX acts before legacy systems detect the event

HPC & Simulation

High-Performance Computing

Sustained peak throughput. No thermal ceiling.

Scientific simulation and HPC workloads demand sustained peak CPU performance over hours or days. Thermal throttling doesn't just slow these jobs — it introduces computational variance that can invalidate simulation results. CoolX provides the thermal stability that HPC workloads require.

The Challenge

A national laboratory running climate simulation workloads was experiencing 8–12% performance degradation on multi-day jobs as processors thermally throttled under sustained load. The degradation was non-uniform across nodes, introducing variance in distributed computation.

The Outcome

CoolX deployment across the simulation cluster eliminated thermal variance entirely. Sustained throughput improved by 11% on average. Multi-day jobs now complete with consistent per-node performance, eliminating the need for post-hoc variance correction in simulation outputs.

Results

11%Throughput improvement

0%Thermal variance

12%Previous degradation

100%Node consistency

Thermal Event Timeline

Legacy: throttle detected

Legacy: recovery time

CoolX: forecast horizon

CoolX: event neutralized

Relative time → CoolX acts before legacy systems detect the event

Built for the workloads that can't throttle.

GPU Cluster Density at Scale

Inference at Hyperscale

PUE Reduction at Hyperscale

High-Performance Computing

Numbers from production deployments.

Which workload is yours?