Lechuck Park

The provided image is a table categorizing the dominant interfaces and protocols used by various infrastructure equipment within a data center to transmit and receive data. The data sources can be explained by dividing them into four main functional tiers:

• IT & AI Core Infrastructure (GPU / server):
  • This is the core source for AI racks, generating the largest point count of data.
  • It predominantly uses Ethernet (dedicated BMC network). The legacy IPMI protocol is actively being retired, making way for the modern Redfish protocol.
• Cooling Systems (CDU, Rear-door HX, Chiller, etc.):
  • CDUs (Cooling Distribution Units) managing Direct Liquid Cooling (DLC) loops handle high-frequency flow, pressure differentials ($\Delta$P), and supply temperatures. They primarily use Ethernet-based Modbus TCP or BACnet/IP.
  • Chillers and cooling towers are currently undergoing a migration to the more secure BACnet/SC (TLS).
• Power Infrastructure (UPS, Smart PDU, Switchgear, BESS, etc.):
  • While UPS and PDUs rely on Ethernet, they internally or optionally retain serial buses like RS-485. They utilize a mix of protocols including SNMPv3, Modbus TCP, and Redfish.
  • High-voltage switchgear and gensets require highly synchronized timing (SNTP/PTP) and utilize utility-grade protocols like IEC 61850 and DNP3.
• Environment & Safety (Rack environmental, Safety chain):
  • Environmental retrofit sensors often utilize RS-485 or LoRaWAN to maintain IP economics.
  • Critical safety chains (EPO, leak detection, fire) deliberately use out-of-band, hardwired DI/DO (dry contacts) to ensure foolproof operation independent of network status.

💡 Direction for Data Collection Improvement in AI Data Centers (AI DC)

While traditional data center protocols are tuned for standard monitoring intervals, the AI DC environment—characterized by high-density GPU servers and liquid cooling architectures—mandates a shift toward ultra-precise and ultra-high-frequency data collection.

AI workloads create massive and sudden spikes in power draw and thermal output. Even a brief delay (Time Constant) in data collection can lead to thermal runaway. Therefore, to ensure practical and stable operations, edge gateways must be upgraded and polling mechanisms for Redfish and Modbus TCP must be optimized to collect telemetry (flow rates, temperatures, power data) between CDUs and GPU servers at millisecond-level frequencies.

📝 Summary

This table illustrates that while Ethernet has become the dominant physical interface across data center infrastructure, the application-layer protocols (Redfish, Modbus, BACnet, hardwired connections) remain highly fragmented depending on the equipment’s specific function (power, cooling, safety). As facilities evolve into AI Data Centers, driving standardization and securing the capability for real-time, ultra-precise data collection—especially around GPUs and liquid cooling systems—has emerged as an urgent operational necessity.This table illustrates that while Ethernet has become the dominant physical interface across data center infrastructure, the application-layer protocols (Redfish, Modbus, BACnet, hardwired connections) remain highly fragmented depending on the equipment’s specific function (power, cooling, safety). As facilities evolve into AI Data Centers, driving standardization and securing the capability for real-time, ultra-precise data collection—especially around GPUs and liquid cooling systems—has emerged as an urgent operational necessity.

#DataCenter #AIDC #Protocols #HighFrequencyTelemetry #LiquidCooling #Redfish #Modbus #InfrastructureOptimization

With Gemini