On December 3, 2025 (Beijing Time), at the highly anticipated re:Invent conference, Amazon AWS made a groundbreaking announcement: its latest self-developed AI chip, Trainium3, has officially hit the market.

Leveraging the advanced 3nm manufacturing process, this new chip represents a quantum leap in performance. When pitted against its predecessor, Trainium3 boasts an impressive 4.4-fold increase in computational performance. This means it can handle complex AI tasks with lightning speed, significantly accelerating the training process for AI models.

In terms of energy efficiency, Trainium3 shines even brighter. It delivers four times the energy efficiency of its forerunner. This is a crucial advantage, especially in today's world where energy conservation and sustainability are paramount. By consuming less power while achieving higher performance, it not only reduces operational costs but also aligns with the global push for greener technologies.

Memory bandwidth is another area where Trainium3 excels. It offers nearly four times the memory bandwidth of the previous generation. This enhanced bandwidth ensures seamless data flow, enabling AI models to access and process large amounts of data more efficiently.

The Trn3 UltraServer is a testament to Trainium3's scalability. A single unit can integrate 144 of these powerful chips, resulting in a staggering total computing power of 362 PFLOPs (petaflops). Moreover, the cluster size can be expanded to accommodate up to 1 million chips, providing unparalleled flexibility and scalability for enterprises with varying AI needs.

Customer tests have yielded remarkable results. The Trainium3 chip has proven to reduce the costs associated with AI model training and inference by up to 50%. This cost reduction is a game-changer for businesses, as it allows them to invest more resources in other critical areas, such as research and development, and accelerates the deployment of AI solutions in various industries.