Introduction: The energy crisis behind the explosion of AI computing power

The energy consumption of the AI training model doubles every 3-4 months, and GPT-4 consumes more than 1,000 megawatt-hours of electricity for a single training session, equivalent to the annual electricity consumption of 3,000 households. When Moore's Law slows down, increasing the density of computing power needs to simultaneously solve the problem of energy consumption and heat dissipation - this is the strategic opportunity for silicon carbide (SiC) semiconductors.

First, SiC technical characteristics: the underlying support of AI high energy efficiency system

Ultra-low loss power conversion

The switching loss of SiC devices (such as MOSFET) is 70% lower than that of silica-based IGBTs, and the power conversion efficiency is increased to 99%+, directly reducing the energy waste of AI server power modules by 30%.

Application Case: Google TPU v4 uses SiC power supply solution, single cabinet power supply density increased to 50kW, support kilocalorie AI training cluster.

High temperature and high frequency stable operation

Temperature resistance of more than 200 ° C, support chip-level liquid cooling system simplified design, help AI server to maintain peak computing power at 80 ° C ambient temperature.

Technology breakthrough: The NVIDIA DGX H100 is equipped with a SiC driver IC, which increases the GPU power frequency to 2MHz and reduces the size by 40%.

Miniaturization and high power density

At the same power, the volume of the SiC module is only 1/5 of the silicon device, freeing up more space for edge AI devices (such as autonomous ECUs, drones) to deploy computing units.

Second, the key landing scene of SiC in the AI industry chain

1. Data center: Full chain innovation from power supply to heat dissipation

Intelligent Power Management:

Delta Electronics introduces 3.6kW/in³ SiC PSU, bringing the power consumption of single-stand AI servers to more than 100kW, while the PUE (energy use efficiency) is reduced to below 1.1.

Liquid cooling system efficiency:

The SiC inverter drives a two-phase immersion cooling pump that uses 45% less energy than conventional air-cooled systems, and Microsoft Azure has deployed this solution in GPT-4 dedicated clusters.

2. Edge Computing: Make AI devices "light"

Autonomous driving domain controller:

The Tesla HW4.0 adopts ST SiC MOSFET, which increases the 12V-48V DC/DC conversion efficiency to 98%, and supports continuous peak operation of FSD chip.

Drones and Robots:

Dji Matrice 350 integrated ROHM SiC power module, battery life extended by 25%, to achieve real-time SLAM (synchronous positioning and mapping) computing power.

3. AI chip manufacturing: the wafer-level thermal management revolution

Nanoscale heat sink materials:

Wolfspeed cooperated with ASML to develop a SIC-based EUV lithography thermal control system, ensuring that the production yield of 3nm AI chips increased to 85%+.

Wafer cutting optimization:

Coherent's laser stealth cutting technology, combined with SiC's high thermal conductivity, reduces the fragmentation rate of AI chip wafers such as Cerebras WSE-3 to 0.01%.

Third, Data perspective: symbiotic growth curve of SiC and AI

Market demand:

Yole predicts that the AI-related SiC device market will reach a CAGR of 62% from 2023 to 2028, and the scale will exceed $4.7 billion in 2028.

Cost inflection point:

With the 6-inch SiC substrate yield increasing to 80%, the cost of SiC MOSFETs in 2025 is expected to be 15% lower than silicon based solutions, triggering a large-scale replacement wave of AI hardware.

Technology Roadmap:

Around 2026, 1200V SiC MOSFETs will fully replace silicon-based devices and become the standard AI server power supply. In 2030, the SiC penetration rate of car-level AI chip power supply system will reach 90%.

4. Challenges and Game Breakers: The path to SiC in the AI era

Material defect control:

The microtubule density needs to be reduced from the current 1/cm² to 0.1/cm² to meet the 10-year life requirement of the AI chip power supply module (the current standard is 5 years).

3D integration technology:

The TSV (silicon through hole) compatible SiC-IC integration scheme is developed to realize the vertical power supply network of AI chip.

Ecological collaborative innovation:

Infineon and TSMC launch the "SiC+3D Package" reference design, enabling AI acceleration card power density to exceed 5kW/cm³.

Conclusion: Superconductor evolution of AI and SiC

As AI evolves toward AGI (Artificial General Intelligence), energy efficiency will become the "Einstein equation" that limits breakthroughs. By reconstructing the power infrastructure, silicon carbide semiconductors are creating "zero loss" energy transmission channels for AI - this is not only the iteration of technology, but also the key lever for intelligent civilization to cross the critical point.