At the 2025 AWS re:Invent conference, Amazon Web Services (AWS) officially announced the release of its latest generation AI training chip - Trainium3. This chip has achieved significant improvements in multiple aspects, bringing new possibilities to AI model training.
According to AWS, the Trainium3 chip uses advanced 3-nanometer process technology, with performance four times that of the previous generation, memory capacity also increased by four times, and power efficiency improved by 40%. These series of improvements enable Trainium3 to better meet the needs of high-load inference and AI training.
The newly released Trainium3UltraServer system can support the construction of large-scale clusters, connecting up to 1 million Trainium3 chips, a number ten times that of the previous generation system. Each UltraServer can accommodate up to 144 chips, greatly enhancing data processing capabilities. Amazon stated that this system will help customers significantly reduce inference costs when using AI cloud services, further promoting the application of AI technology.
In addition, AWS revealed plans for the development of Trainium4, stating that this new-generation chip will support collaboration with NVIDIA GPUs. By supporting NVIDIA's NVLink Fusion high-speed interconnect technology, Trainium4 not only can expand overall performance but also utilize Amazon's self-developed low-cost server architecture. This move may lower the barrier for migrating existing AI applications optimized for NVIDIA to the Amazon cloud platform.
Key Points:
🌟 Trainium3 performance is four times better, with memory and power efficiency increased by four times and 40% respectively.
🔗 The UltraServer system can support 1 million chips, enabling ultra-large-scale clusters.
🚀 Trainium4 will support collaboration with NVIDIA GPUs, lowering the barrier for AI application migration.
