The number of GPUs per rack continues to increase, while the number of dies is growing at an even faster rate.
Total rack power consumption is expected to reach 1MW, presenting three critical challenges:

high-voltage DC-DC conversion in a smaller form factor;
higher capacitance density and a faster di/dt response for GPU core power;
and the placement of embedded MLCCs among numerous PTHs (Plated Through-Holes) while maintaining power integrity.
Samsung Electro-Mechanics introduces capacitor solutions for 1MW racks.

① Overcoming Rack Space Constraints with a High-Voltage DC-DC Converter

In a high-efficiency power system architecture, an HV DC-DC converter can be deployed on the GPU computing board.
In a 1MW-class high-density rack power system, snubber capacitors play a critical role in suppressing spike noise generated by high-speed MOSFET switching circuits. Spike noise adversely affects the power integrity (PI) of the power system and places stress on power devices such as MOSFETs, potentially compromising reliability. Ceramic capacitors have significantly lower ESL than film capacitors and therefore provide outstanding response characteristics, rapidly absorbing ultra-high-frequency parasitic harmonics and sudden surge currents. The reduction in effective capacitance caused by DC bias under high voltage can be sufficiently mitigated by selecting an appropriate dielectric, such as C0G or X7R, and connecting capacitors in parallel. Ceramic capacitors also offer higher capacitance density per unit area in space-constrained designs. Consequently, compact ceramic capacitors with superior ESL performance are key components for maximizing the power density of high-voltage DC-DC converters, minimizing switching losses, and ensuring stable operation of 1MW data center rack power systems.

Ceramic Film Capacitor ESL/ESR ●●● ● Capacitance Density ●●● ●   Capacitance ● ●●● Size Thickness ●●● (Small and thin) ● (Large and thick) Application High frequency Noise decoupling Power smoothing AC filter

CL55C543JI6MPN Measurement Graph

② Proposal for Compact, High-Capacitance MLCCs for GPU Core Power

The MLCC capacitance density required for GPUs at least doubles each year. GPU power consumption is increasing much faster than the area of the GPU package allocated to core power.
This means that more capacitors are required even as the available space remains limited.
When VPD(Vertical Power Delivery) is applied, the space available for MLCC placement becomes even more constrained.
The following simulation results explore how capacitance density can be maximized.

"AI MLCC's Law": The MLCC capacitance required per GPU doubles every year.

Maximize cap density vs Impedance

Total MLCC count and capacitance available in 100x100 core power area

	MLCC	Usage (pcs)	Capacitance value(uF)	Density	ESL
	0201 10uF	19,000	190,000	●	●●●
	0402 47uF	7,900	371,300	●●●	●●
	0603 100uF	4,200	420,000	●●●●	●

Placing 0603 100㎌ components in a 100x100 area provides the highest capacitance density, but the ESL is too high.
By contrast, 0201 10㎌ components offer much better ESL performance but insufficient capacitance density.
The 0402 47㎌ option provides the optimal balance between capacitance density and ESL.

Optimal points

Modern inference servers require higher di/dt, so the number of 0201 10㎌ components was increased to reduce impedance. This is because 0201 10㎌ components inherently have low ESL.

To supplement capacitance, high-capacitance-density 0603 100㎌ components were also added. Although total capacitance decreased, impedance characteristics improved. Ultimately, an optimized design must account for the trade-off between capacitance provided by 0603 100㎌ components and impedance performance provided by 0201 10㎌ components. When equal numbers of the three capacitor types were used, impedance actually worsened, demonstrating that the mixing ratio is critically important.

Another challenge is that the PTH pitch is becoming increasingly narrow to deliver more power to the GPU. Thus, it has become more important to increase capacitance density in compact packages such as 0201 and 0402. Using a capacitor larger than the PTH pitch requires additional PCB routing, which increases ESL.

As noted above, inference servers require faster di/dt, making it essential to secure high capacitance while maintaining low ESL in smaller package sizes.

③ Embedded MLCC Technology

Embedding capacitors within the PCB is a highly effective solution for increasing capacitance density while reducing ESL (equivalent series inductance).

Initially, vertically embedding MLCCs among multiple PTHs (plated through-holes) was considered, but the 2 mm PTH pitch was too wide.
A horizontal-mounting approach was then considered:
power is supplied through the MLCC's external electrodes, eliminating the necessity for PTHs. This enables a narrower power-supply pitch and enhanced electrical performance.

Horizontal vs Vertical Embedding

Please click here for sample requests or any other inquiries.
USA (semai.newsletter@samsung.com)
EU (semcoeurope@samsung.com)
SEA (sempl.automlcc@samsung.com)
CHN (semco.automlcc@samsung.com)
JPN (semj.automlcc@samsung.com)
KOR (automotive.sem@samsung.com)