Abstract​

AI servers and data centers require high-performance computing and high-speed communications in order to process and analyze large volumes of data. Under these circumstances, numerous high-performance ICs (CPU, GPU, ASIC, etc.) are applied. This leads to heat dissipation, which causes large current transients during intensive computation. Ultra-high-capacitance MLCCs are suitable for decoupling high-frequency noises and responding to peak current demands.

Samsung Electro-Mechanics presents the ultra-high-capacitance MLCCs applicable to various AI server applications: CL31X227MRKNNW# (1206 inch, 220 µF, 4Vdc, X6S) and CL32X337MSVN4S# (1210 inch, 330 µF, 2.5 Vdc, X6S).

① Application of High-Temperature / Ultra-High-Capacitance MLCCs, CPO Switch (Co-Packaged Optics Switch)​

A CPO Switch integrates conventional optical components and network switch chips into a single package. This enables ultra-high-speed data transmission in data centers and high-performance computing environments. The addition of optical engines increases temperature and power consumption, thus requiring ultra-high-capacitance, high-temperature MLCCs of X7 or higher.

(Suggestions : CL32X337MSVN4S#, CL31X227MRKNNW#, CL32Z227MSVN4S#, CL31Z107MRKN4N#)

② Proposal: Thickness-Constrained PCIe Card / OAM Socket, 220 µF Ultra-High-Capacitance MLCCs​

A PCIe Card (Peripheral Component Interconnect Express Card) supports high-speed data transfer for the fast transmission of large-volumes of data. Graphics cards, storage devices, network adapters, AI accelerators and other components require high-power and high-speed data handling. To ensure power stability, multiple high-capacitance MLCCs are demanded.​

When multiple PCIe cards are installed within an AI server (refer to Figure 2), space constraints and thermal considerations make thickness-limited ultra-high-capacitance MLCCs desirable. We propose ultra-high-capacitance MLCCs (thickness ≤ 2.0mm), such as CL31X227MRKNNW# and CL31A227MQKNNW#.​

OAM Socket (Open-Compute-Project Accelerator Module Socket) is an interface designed for high-performance modules including GPUs for AI/ML workloads or custom ASICs. Using a custom ASIC in an Accelerator Module enables optimization of performance-per-watt and the removal of unnecessary circuitry to place memory closer. However, custom ASICs contain a large number of switching circuits that generate rapid transient currents. To handle such sudden current changes, high-capacitance MLCCs are required. They also serve to decouple, remove high-frequency noise and stabilize voltage.

As shown in Figure 3, numerous ultra-high-capacitance MLCCs can be mounted on the OAM PCB Land Side — e.g., CL31X227MRKNNW# (1206 inch / 1.9 t, 220 µF, 4 Vdc, X6S) and CL31A227MQKNNW# (1206 inch / 1.9 t, 220 µF, 6.3 Vdc, X5R) — contributing to improved space efficiency.​

Samsung Electro-Mechanics is mass producing ultra-high-capacitance MLCCs (1206–1210 inch, 100–330 µF) which offer the high quality reliability and stable electrical performance demanded by AI servers and data centers.