引導式測驗 · 中等 等級

AI 資料中心的電力和冷卻測驗

解釋密集人工智慧加速器系統的資料中心供電、散熱、直接液體冷卻、PUE 和水權衡。

問題 1 的 8

Why can AI accelerator racks require different cooling designs than lower-density server racks?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Why can AI accelerator racks require different cooling designs than lower-density server racks?
    • Because accelerators produce less heat than most storage equipment
    • Because rack dimensions eliminate the facility heat load
    • They can concentrate more electrical power and heat in a smaller space
    • Because accelerator racks do not need electrical distribution
  2. What does direct-to-chip liquid cooling do?
    • Circulates coolant through the entire server room instead of using air
    • Circulates coolant near hot components to carry heat into a facility loop
    • Removes all heat without a heat-rejection system
    • Exhausts server air directly outside without heat exchange
  3. What does PUE compare?
    • IT equipment energy divided by total facility energy
    • Water consumed per IT energy unit
    • Compute output per rack area
    • Total facility energy divided by IT equipment energy
  4. What does PUE not measure by itself?
    • Water consumption or carbon intensity
    • Total facility energy
    • IT equipment energy
    • The energy ratio between facility and IT equipment
  5. Why can cooling designs trade electricity against water use?
    • Lower cooling energy always means lower water use in every climate
    • Liquid cooling eliminates all facility water requirements
    • Cooling methods such as evaporation have different energy and water requirements
    • Power draw has no relationship to heat removal
  6. Which examples are part of a data center’s cooling load?
    • GPU compute cores and memory banks
    • Fans, pumps, chillers, and cooling towers
    • Training jobs and optimizer states
    • Storage systems and network interfaces
  7. What monitoring can help detect liquid-cooling risks?
    • Model validation accuracy and label quality
    • Batch size and data split consistency
    • The number of cloud regions used for training
    • Coolant flow, pressure, temperature, and leak sensors
  8. Why should facility planning use measured workload profiles?
    • Measured power and utilization can differ from nameplate maximums
    • GPU specifications fully determine site impact
    • A single benchmark predicts facility demand for every workload
    • Local climate has no effect on cooling design