AI数据中心推动对SiC、GaN功率半导体的需求
人工智能（AI）数据中心的普及带动电力消耗上升，凸显碳化硅（SiC）与氮化镓（GaN）等化合物功率半导体的重要性。
原文深度解读
本文主题为AI数据中心电力消耗上升推动碳化硅（SiC）与氮化镓（GaN）等化合物功率半导体需求。涉及主体包括韩国电子通信研究院（ETRI）高级首席研究员Bae Sung-beom、韩国功率器件产业协会、英伟达（NVIDIA）800V直流数据中心供电架构。时间为10月7日，地点为釜山农心酒店。关键数字包括800V直流电压、外部电网数千至数万伏电压。与电子元器件供应链的关系在于：数据中心高压供电架构升级将影响SiC和GaN功率半导体的应用阶段、衬底材料与器件结构，并涉及可靠性问题对商业化的制约。
- AI数据中心普及带动电力消耗上升，数据中心正采用高压供电网络以更高效地输送电力，而传统硅（Si）功率半导体在降低功率损耗方面存在局限。
- ETRI高级首席研究员Bae Sung-beom于10月7日在釜山韩国功率器件产业协会主办的功率半导体产学研K论坛上表示，AI产业发展正在全球范围内造成电力短缺，由于电力基础设施难以按实际需求增长速度扩张，高效利用有限能源日益重要。
- 数据中心全天候运行GPU等大规模计算资源，随着AI模型增大，所需GPU和存储容量增加，推高电力消耗；如何以最小损耗将外部电网电力输送到服务器也日益重要。
- 英伟达800V直流数据中心供电架构在此背景下出现，该系统以高于传统系统的800V直流电压在数据中心内部配电，简化现有供电架构并以更高电压输电，减少电力转换和布线损耗。
- 随着供电电压升高，功率半导体作用更重要；高压大功率环境下，设计耐受更高电压的器件会更难降低功率损耗和发热，因此SiC和GaN等化合物功率半导体被视为替代方案，它们可在比Si更高电压和温度下工作，减少转换损耗，并因高速开关能力帮助提升转换设备效率、实现更小元件。
- 应用因电压水平和转换阶段而异：SiC适合高压大功率环境，可用于将外部电网数千至数万伏电力转换为数据中心内部可用电力；GaN在相对较低电压下提供快速开关性能，可帮助降低服务器内部降压至GPU等设备所需电压时的损耗。商业化仍需解决可靠性问题，不同材料热膨胀系数差异可能导致晶圆翘曲或开裂，衬底热导率也影响器件输出和效率，业界正研究新衬底材料和垂直器件结构。
供应链影响
- 若AI数据中心持续采用800V直流等高压供电架构，SiC功率器件在外部电网到数据中心内部的转换环节可能获得更多应用机会，具体取决于高压架构的落地节奏和器件可靠性进展。
- GaN器件可能因服务器内部降压至GPU所需电压的需求而受益，但其应用规模可能受限于较低电压场景的渗透速度以及可靠性问题的解决程度。
- 化合物半导体制造中晶圆翘曲或开裂、衬底热导率等可靠性问题，可能影响SiC和GaN器件的良率与商业化进度，进而可能对相关功率半导体供应稳定性产生间接影响。
- 业界研究新衬底材料和垂直器件结构，可能推动上游衬底材料与器件结构相关供应链环节发生变化，但实际影响取决于研发成果的产业化时间。
- AI数据中心电力基础设施扩张速度若跟不上需求增长，可能促使数据中心更积极采用高效功率转换方案，从而间接影响SiC和GaN功率半导体在数据中心供电链条中的需求结构。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
Bae Sung-beom，韩国电子通信研究院（ETRI）高级首席研究员，于10月7日在釜山东莱区农心酒店举行的、由韩国功率器件产业协会主办的功率半导体产学研K论坛上发表演讲。（照片：THE ELEC）
随着人工智能（AI）数据中心的普及，电力消耗正在上升，凸显了碳化硅（SiC）和氮化镓（GaN）等化合物功率半导体的重要性。数据中心正在采用高压供电网络，以更高效地输送更多电力。传统硅（Si）功率半导体在降低功率损耗方面存在局限。
韩国电子通信研究院（ETRI）高级首席研究员Bae Sung-beom于10月7日在釜山农心酒店举行的韩国功率器件产业协会功率半导体产学研K论坛上表示，AI产业的增长正在全球范围内造成电力短缺。“由于电力基础设施难以按实际需求增长的速度扩张，高效利用有限能源正变得越来越重要，”Bae说。
数据中心全天候运行图形处理器（GPU）等大规模计算资源。随着AI模型变得更大，数据中心所需的GPU数量和存储容量增加，推高了电力消耗。如何以最小损耗将外部电网供应的电力输送到服务器也正变得越来越重要。
提高电力效率需要减少将外部电力转换为服务器或GPU可处理电压时产生的损耗。英伟达（NVIDIA）的800伏直流（DC）数据中心供电架构正是在这一背景下出现的。800V直流系统以高于传统系统的800伏直流电压在数据中心内部配电。它简化了现有供电架构，并以更高电压输电，减少了电力转换和布线过程中的损耗。
随着供电电压升高，功率半导体的作用变得更加重要。在高压大功率环境中提高电力效率存在局限，因为设计耐受更高电压的器件会使降低功率损耗和发热变得更加困难。
SiC和GaN等化合物功率半导体正被视为替代方案。它们可以在比Si更高的电压和温度下工作，并减少电力转换过程中的损耗。它们的高速开关能力也有助于提高电力转换设备的效率，并使元件更小。
应用因电压水平和电力转换阶段而异。SiC适合高压大功率环境。它可用于将外部电网供应的数千至数万伏电力转换为数据中心内部可用电力的设备。GaN在相对较低的电压水平下提供快速开关性能。它可以帮助减少服务器内部将电压降至GPU和其他计算设备所需水平时的电力损耗。
商业化仍需解决可靠性问题。化合物半导体通过组合不同材料来形成器件。材料之间热膨胀系数的差异可能导致晶圆翘曲或开裂。衬底的热导率也会影响器件输出和效率。业界正在研究使用新衬底材料和垂直器件结构来解决这些问题的方法。
Bae Sung-beom, senior principal researcher at the Electronics and Telecommunications Research Institute (ETRI), speaks at the Power Semiconductor Industry-Academia-Research K-Forum hosted by the Korea Power Device Industry Association at Hotel Nongshim in Dongnae, Busan, on Oct. 7. (Photo: THE ELEC)
Power consumption is rising with the spread of artificial intelligence (AI) data centers, highlighting the importance of compound power semiconductors such as silicon carbide (SiC) and gallium nitride (GaN). Data centers are adopting high-voltage power networks to deliver more electricity efficiently. Conventional silicon (Si) power semiconductors have limitations in reducing power losses.
Bae Sung-beom, a senior principal researcher at the Electronics and Telecommunications Research Institute (ETRI), said Oct. 7 at the Korea Power Device Industry Association's Power Semiconductor Industry-Academia-Research K-Forum at Hotel Nongshim in Busan that the growth of the AI industry is causing power shortages worldwide. “Because it is difficult to expand power infrastructure at the pace of actual demand growth, it is becoming increasingly important to use limited energy efficiently,” Bae said.
Data centers operate large-scale computing resources such as graphics processing units (GPUs) around the clock. As AI models grow larger, the amount of GPUs and storage capacity required by data centers increases, driving up power consumption. How efficiently electricity supplied from external power grids can be delivered to servers with minimal losses is also becoming increasingly important.
Improving power efficiency requires reducing losses that occur when electricity from outside is converted to voltages that servers or GPUs can handle. NVIDIA's 800-volt direct-current (DC) data center power architecture has emerged against this backdrop. The 800V DC system distributes electricity inside data centers at a higher 800-volt DC voltage than conventional systems. It simplifies the existing power-supply architecture and transmits electricity at higher voltage, reducing losses during power conversion and wiring.
As power-supply voltages rise, the role of power semiconductors becomes more important. Improving power efficiency in high-voltage and high-power environments has limitations because designing devices to withstand higher voltages makes it more difficult to reduce power losses and heat generation.
Compound power semiconductors such as SiC and GaN are being considered as alternatives. They can operate at higher voltages and temperatures than Si and reduce losses during power conversion. Their ability to switch at high speeds also helps improve the efficiency of power-conversion equipment and enables smaller components.
Applications vary depending on voltage levels and power-conversion stages. SiC is suited to high-voltage, high-power environments. It can be used in equipment that converts electricity supplied from external power grids at thousands to tens of thousands of volts into power that can be used inside data centers. GaN offers fast switching performance at relatively lower voltage levels. It can help reduce power losses when voltage is stepped down inside servers to levels required by GPUs and other computing devices.
Reliability issues need to be addressed for commercialization. Compound semiconductors are manufactured by combining different materials to form devices. Differences in the coefficients of thermal expansion between materials can cause wafers to warp or crack. The thermal conductivity of substrates also affects device output and efficiency. The industry is researching ways to address these issues using new substrate materials and vertical device structures.