Huayuan Securities has released a research report highlighting the launch of Alphabet's eighth-generation Tensor Processing Unit (TPU), which includes the TPU 8t, specialized for training, and the TPU 8i, designed for inference. Compared to the previous generation, the new chips offer a 2.7x improvement in training performance per dollar and an 80% increase in inference performance per dollar. The new processors incorporate fourth-generation liquid cooling technology, doubling performance per watt. Additionally, the integration of diamond substrate cooling technology is expected to further enhance thermal management efficiency. The main points from Huayuan Securities are as follows.
Alphabet's eighth-generation TPU demonstrates a significant performance leap over the seventh-generation Ironwood model. On April 22, 2026, Alphabet officially introduced the eighth-generation TPU at its Cloud Next event, comprising the TPU 8t and TPU 8i. The TPU 8t is optimized for large-scale, computationally intensive training workloads, delivering greater computational throughput and enhanced scalable bandwidth. The TPU 8i features higher memory bandwidth and is specifically designed for latency-sensitive inference tasks. In terms of training cost-efficiency, the TPU 8t provides a 2.7x improvement over the Ironwood TPU, making it suitable for massive training projects. For inference cost-efficiency, the TPU 8i offers an 80% gain over its predecessor, particularly for achieving low latency in large Mixture-of-Experts (MoE) models.
The eighth-generation TPU continues to optimize efficiency with its fourth-generation liquid cooling design, doubling the performance per watt for both chips. To address power consumption challenges, Alphabet has enhanced overall stack efficiency and integrated power management features that dynamically adjust power usage based on real-time demand. Beyond chip-level optimizations, Alphabet has made system-level architectural improvements from the chip to the data center. For instance, integrating network connectivity directly onto the same chip significantly reduces energy consumption for data transfer between TPUs. Through continuous innovation in both hardware and software, Alphabet's data centers now deliver six times more computational power per unit of electricity compared to five years ago. Furthermore, the TPU 8t and TPU 8i utilize fourth-generation liquid cooling, enabling performance densities unattainable with air cooling.
The integration of diamond as a substrate material promises further advances in thermal dissipation. A recent technical report from Diamond Foundry indicates that single-crystal diamond has a thermal conductivity 16 times greater than silicon and approximately six times greater than materials like silicon carbide, copper, aluminum nitride, silver, and gold. Diamond serves not only as a heat conduction medium but also, when integrated as a substrate in high-power chip packaging, provides exceptional cooling performance. By redistributing overall thermal resistance and the temperature drop across the substrate, it enables two-phase cooling at the chip level. This design can enhance thermal dissipation performance by 10 to 100 times.
Regarding investment recommendations, liquid cooling technology is seeing increasing adoption with multiple technical approaches advancing in parallel. The report suggests focusing on companies involved in liquid cooling technologies across various segments: 1) Full systems: include firms like Invic and Shenling Environment. 2) Cold plates: companies such as Yidong Electron, Kechuang Xinyuan, Siquan New Material, Shuobede, Hongfuhan, Tongfei Co., Jiebang Technology, and Far East Holdings. 3) CDUs and other components: firms like Xingrui Technology, Gaolan Co., Feilong Co., Chuanhuan Technology, Chuannrun Co., Yimikang, Zhongshi Technology, Shuguang Shuchuang, and Dayuan Pump Industry. 4) Coolants: suppliers including Runhe Material, Yongtai Technology, Juhua Co., Dongyangguang, and Xinzhoubang. 5) Thermal interface materials (TIM): companies such as Feirongda and Debang Technology. 6) Diamond cooling: firms like Sifangda, Woerde, Huanghe Xuanfeng, Liliang Diamond, Guoji Jinggong, and Huifeng Diamond.
Risk factors include potential delays in technological progress, lower-than-expected capital expenditure overseas, and intensifying market competition.
Comments