Hong Kong - On December 5, 2024, Amazon Web Services (AWS) announced the availability of Amazon Elastic Compute Cloud (Amazon EC2) Trn2 instances, powered by the new Amazon Trainium2 AI chip, during the 2024 re:Invent global conference. The Trn2 instances promise a 30-40% cost-performance improvement over existing GPU-based EC2 instances, facilitating the training and deployment of advanced AI models.
David Brown, AWS Vice President of Compute and Networking, stated that the Amazon Trainium2 is designed to support large-scale generative AI workloads, addressing the need for efficient training and inference as model parameters approach trillions. The new EC2 Trn2 UltraServers offer high-speed training and inference capabilities, enabling businesses to deploy large-scale models more rapidly and cost-effectively.
The Trn2 instances integrate 16 Trainium2 chips, delivering up to 20.8 Petaflops of floating-point performance, ideal for training billion-parameter models. The UltraServers allow customers to scale beyond the limitations of individual Trn2 instances, enhancing training speed and market readiness.
AWS is collaborating with Anthropic to develop Project Rainier, a distributed computing cluster utilizing Trn2 UltraServers, aimed at training AI models with unprecedented computational power. This initiative is expected to yield the largest available AI computing cluster globally.
Additionally, AWS announced plans for the next-generation Amazon Trainium3 chip, set to launch by the end of 2025, which will further enhance performance and energy efficiency for advanced AI workloads.