Amazon Last night, during the AWS re:Inve 2025 eve, a new generation of proprietary chips for artificial ielligence processing named Trainium3 iroduced that it is supposed to Training speed of models to increase significaly.
According to Amazon, the Trainium3 chip offers four times better performance and four times more memory than its predecessor. This improveme makes progress both in the field of training and derivation of artificial ielligence models.
Optimum energy consumption is also one of the main axes of the new generation. According to Amazon, Trainium3 systems About 40 perce They consume less energy and at the same time provide more computing power. Given the rapid growth of data ceers and increasing energy demand, AWS says the new design can reduce infrastructure strain and lower operational costs associated with AI processing.
Amazon also has a huge system called Trainium3 UltraServer It has revealed that it uses 3nm Trainium3 chips. Each UltraServer includes 144 Trainium3 chips, and customers can connect thousands of these systems together.
At the largest possible scale, a single server unit can One million Trainium3 chips to use; A figure that shows a tenfold improveme compared to the previous generation.
Amazon’s first Trainium3 chip customers
Several companies have already started using Trainium3. AhropicJapanese startup Karakuriservice SplashMusic and company Descartes Among the customers are those who, according to Amazon, have achieved better performance in inference and other advaages by benefiting from this new chip.
The company also confirmed that it has started developme of its next-generation proprietary processor, Trainium4. The most importa change meioned for the next generation is support for NVIDIA NVLink Fusion which enables the construction of massive and iegrated artificial ielligence processing clusters.
This choice shows that Amazon no longer iends to use its chips as a complete replaceme for processors Nvidia to iroduce, but rather to build a hybrid ecosystem where AWS’s proprietary chip can work alongside and even compleme Nvidia’s GPUs.
