Function of capacitor in power supply and signal link of AI server
In the AI server, the dynamic change of the working load is different from the general computing: accelerator (such as GPU, TPU, ASIC) quickly switches the power state, memory module needs strict voltage regulation in a wide frequency range, and peripheral systems (storage, network) may have a large current transient due to the change of I/O load. The capacitors in these systems have several basic functions:
High energy storage/smoothing
Cross-band decoupling/bypass
Transient response
Power failure protection
Table 1: Classification of capacitor characteristics for AI servers. Picture source: Ymin)
Table 1 lists the breakdown of the most critical specifications and their correspondence to the AI server requirements:
Capacitor selection by AI server subsystem
Motherboard and VRM level:
Challenges: fast transient current, tight voltage regulation, switching noise
Best capacitor type: multilayer polymer solid aluminum, conductive polymer tantalum, standard solid polymer
Design criteria: Hybrid bulk and high frequency decoupling (polymer+MLCC) to minimise inductance, derated design
Power supply (AC/DC, DC/DC converter):
Challenges: large ripple, low conversion efficiency, long operating time
Best capacitor type: wet electrolysis (input), mixed polymer (output), polymer or multilayer (high-frequency filtering)
Design criteria: large capacity input side capacitor, low ESR output, control thermal conditions
Memory/SSD/Power down buffer:
Challenge: Stored energy must be available in the event of power failure
Best capacitor type: wet electrolysis, mixed polymer, multilayer solid polymer
Design principle: calculate E=½ CV ², ensure redundancy, manage leakage and aging
Network/Interconnect/Switch:
Challenges: Burst Flow, EMI, Dynamic Load
Best capacitor type: low ESR solid polymer, multilayer solid polymer
Design principle: use high rated ripple capacitor to minimize parasitic, and use in combination with MLCC
Gateway, aggregation node, external interface:
Challenges: Bridging systems require robust noise suppression.
Best capacitor type: multilayer polymer solid aluminum, polymer/hybrid type
Design principle: broadband decoupling, ripple suppression and thermal derating