Rose-gold DFX-G7 PCIe Gen7 Smart Switch chip engraved with the Datafabrix logo, on a circuit board with glowing PCIe traces.

PCIe Gen7 Smart Switch · in development

Intelligence at the Heartof the PCIe Fabric.

Datafabrix is building an AI-native infrastructure platform: the DFX-G7 PCIe Gen7 Smart Switch, a Smart Backplane with a confidential sensor fabric, and a DCIM software stack that turns their telemetry into predictive maintenance.

128 GT/s per lanePCIe Gen7 signaling
AI on fabric telemetryThermal · fault · load
Predictive, not reactiveMaintenance on evidence

Conceptual illustration

01The Datafabrix stack

Intelligence from silicon to data center.

Four layers, one telemetry model. The backplane senses, the switch sees and infers, DCIM correlates across the fleet, and operators act before anything breaks.

Exploded view of the Datafabrix stack: switch chip, backplane, server rack and data-center hall connected by rose-gold data beams.
Chip → board → rack → hallConceptual illustration
FacilityDatafabrix DCIMCorrelate · Plan · Act
AI modelsPredictive IntelligenceThermal · Fault · Load
SystemSmart BackplaneConfidential sensing
FabricSmart PCIe FabricDeep telemetry
SiliconPCIe Gen7 Smart SwitchDFX-G7 · AI-ready
  • Sense at the source. Per-bay temperature, power, airflow and device health from the Smart Backplane sensor fabric.
  • Infer in the fabric. DFX-G7 combines backplane signals with its own port, queue and link telemetry to score risk in real time.
  • Plan across the fleet. Datafabrix DCIM joins silicon signals with facility data and schedules work before failure.
03What the AI predicts

Three questions, answered before they become incidents.

The DFX-G7 AI engine and DCIM models are being designed around the three failure modes that cost AI data centers the most: heat, hardware faults and imbalanced load.

ThermalA server rack with rose-gold heat haze rising from hotspots.

Where will it run hot?

Bay and component temperatures, fan state and traffic intensity forecast hotspots hours ahead, so cooling adjusts before throttling.

Inputs
Bay temps, airflow, power draw, port utilization
Output
Hotspot forecast with time-to-threshold
FaultAn NVMe drive in a connector with a rose-gold diagnostic ring and a degrading signal trace.

What is about to fail?

Correctable-error clusters, link retrains, latency drift and drive health logs combine into a per-device failure risk score.

Inputs
AER errors, LTSSM events, SMART logs, latency
Output
Risk score, evidence and replacement window
LoadEight GPU modules connected through a central switch with rose-gold traffic paths.

Where should traffic go?

Queue growth and per-port utilization predict congestion, so traffic classes and placements are rebalanced before jobs stall.

Inputs
Queue depth, credits, bandwidth, thermal headroom
Output
QoS reweighting and placement recommendations

Design targets for the Datafabrix platform. Capabilities are in development and not yet validated in production.

04Predictive maintenance

Maintenance on a schedule, not on an outage.

Most data centers still learn about a failing drive, link or fan when it fails. Datafabrix moves that moment earlier, from the incident to the plan.

A technician walking down a data-center aisle toward a rack highlighted for a planned maintenance window.
Planned window · replacement part readyConceptual illustration

Today

Reactive

  • Alarm fires after the drive or link fails
  • Jobs stall, data rebuilds, SLAs slip
  • Technician dispatched without root cause
  • Spares and downtime are unplanned

With Datafabrix

Predictive

  • Risk rises days ahead, with the evidence
  • Workloads drain before the window
  • Work order names the bay, part and cause
  • Downtime is scheduled and short
Sense

Backplane sensor fabric streams bay-level health.

Observe

DFX-G7 adds port, queue and link telemetry.

Predict

On-switch AI scores thermal, fault and load risk.

Plan

DCIM ranks risks and books a maintenance window.

Act

Drain, replace, verify. The model learns from the outcome.

05Where it applies

Built for AI. Relevant wherever PCIe connects.

Primary focus

AI Data Centers

GPU and accelerator servers where PCIe links hosts, accelerators, NICs and local NVMe, and where heat and congestion cost the most.

Target marketView →
Storage

NVMe Storage

JBOF and high-density SSD platforms with per-drive telemetry and predicted replacement windows.

Target marketView →
Cloud · HPC · Enterprise

Composable infrastructure

Multi-host device pools, HPC node I/O and serviceable enterprise platforms.

Adjacent marketsView →

Not just moving data. Making the PCIe fabric intelligent.

We work with hyperscalers, OEMs and silicon partners on Gen7 architecture requirements. Bring us your topology.