亚马逊云科技推出自研AI芯片Amazon Trainium

Core Insights - Amazon Web Services (AWS) announced the launch of the new P6E GB300 series and the Trainium 3-based Trn3 UltraServers at the 2025 re:Invent global conference, emphasizing their commitment to providing top-tier computing power for demanding AI workloads [1][2][3] - The introduction of Amazon AI Factories allows customers to deploy dedicated AWS AI infrastructure within their own data centers, ensuring physical and logical isolation while maintaining access to AWS's advanced AI services [1][3] Product Launches - The P6E GB300 series utilizes NVIDIA's latest GB300 NVL72 system, designed to deliver exceptional reliability and performance for large enterprises, including NVIDIA's own Project Ceiba and organizations like OpenAI [1][3] - AWS's self-developed AI chip, Amazon Trainium, is recognized as one of the best inference systems globally, with deployment speeds significantly faster than previous chips, contributing to a multi-billion dollar business that continues to grow [2][4] Future Developments - The Trainium 3 UltraServers are now officially available, and AWS is actively developing Trainium 4, which is expected to achieve substantial improvements over Trainium 3, including a 6x increase in FP4 computing performance, 4x increase in memory bandwidth, and 2x increase in high-bandwidth memory capacity [2][5]