Immersion Server Demonstrated Reliability at Datacenter Scale Deployment

Author: Austin Hipes (Chief Technologist and VP of Engineering, Unicom Engineering) Collection: 2025-OCP-EMEA-Summit-Immersion-Cooling

Summary

Austin Hipes (Unicom Engineering) presented the most comprehensive publicly disclosed real-world reliability dataset for single-phase immersion cooling at OCP EMEA Summit 2025 (Dublin, 29–30 April). The dataset covers 3,590 servers — 2,932 second-generation Xeon and 658 third-generation Xeon — deployed in GRC ICEraQ FLEX series tanks with PA04 dielectric fluid at a Tier 1 data center in the APJ region. Unicom Engineering has provided all hardware support since the deployment began in September 2022, giving approximately two and a half years of production failure data from a single, controlled environment. The data center also hosts air-cooled and Direct-to-Chip (DTC) servers of the same model, enabling direct cooling technology comparisons.

The headline result is that immersion server uptime equals or exceeds 5 Nines (99.999%+) in both the all-failures and filtered-failures scenarios, and the customer's own measurement is that immersion server downtime is 1/8th of their air-cooled servers of the same model over the same period. The most important analytical finding is that the majority of elevated failure events were caused by poor conversion practices rather than by the immersion environment itself. Nine specific conversion errors were documented at this facility — work performed on data center floors without ESD controls, no torque-limited drivers, tissue-paper TIM cleaning, fan removal from PSUs, and standard non-immersion DAC cables installed face-up with no cable management. Even with all nine errors present, immersion failure rates remained better than air. Failure analysis of returned motherboards confirmed LGA socket damage (CPU reinserted without alignment clips, pins bent), debris from tissue paper in DIMM slots causing memory errors, and one catastrophic case where un-cleaned conductive TIM floated and shorted two CPU VRMs — melting the PCB — with no impact to any other server in the tank.