Data Center Infrastructure for GPU Systems
Data Center covers the physical and operational envelope for GPU systems: power distribution, cooling, rack integration, cabling, redundancy, and monitoring. The exam tests whether you can translate GPU and network capacity into a deployable, serviceable facility.
Core rules
- Plan by rack power density, not average hall density. A rack of eight 10kW GPU servers is an 80kW rack; traditional air cooling may not remove that heat.
- Use liquid cooling, direct-to-chip or immersion, when air cannot meet the load, and verify facility water temperature, flow, pressure, and leak detection.
- Size power for redundancy and derating. A 208V three-phase branch at 80kW draws about 222A at unity power factor; at 0.95 PF it is about 234A. A 250A circuit is not acceptable with normal 80% derating.
- Separate data, management, and storage networks; label both ends of every fiber; respect bend radius and transceiver reach.
- Monitor inlet and outlet temperatures, GPU clocks and throttling, PSU load, fan speed, and leak sensors. Treat thermal throttling as a performance incident.
Worked example
An 80kW rack on 208V three-phase: current = 80,000 / (1.732 x 208 x 0.95) is about 234A. For continuous load at 80% of breaker rating, required breaker = 234 / 0.8 is about 292A, so specify a 300A feed or split the rack across feeds. The same rack on an air-cooled 20kW-per-rack design is infeasible; liquid cooling or a lower-density deployment is required.
Traps
Counting only server nameplate watts and ignoring switch, PSU, and cooling overhead; forgetting UPS/PDU redundancy; using average power for thermal design; and mixing single-mode and multimode optics without checking reach and FEC.