Subsidiary of Amazon.com (AMZN, Financials), Amazon Web Services announced the public availability of its AWS Trainium2-powered Amazon Elastic Compute Cloud instances at the AWS re:Invent conference; AMZN stock is up 1.1% on last look, changing hands for $213.07.
These new GPU-based EC2 instances promise a 30–40% price-performance gain over present ones, according to a statement from Amazon. Along with this release, AWS presented its next-generation Trainium3 AI processors, meant to improve performance and efficiency in future deployments, and Trn2 UltraServers, meant to enable big-scale AI workloads.
Comprising sixteen Trainium2 processors, Trainium2 instances provide 20.8 peak petaflops of computational capability. They are best suited for training and implementing generative artificial intelligence projects requiring great computing efficiency as well as for large language models. Using the ultra-fast NeuronLink interface, the recently added Trn2 UltraServers increase the capabilities of these instances by joining 64 Trainium2 processors, Amazon said. Faster training times and real-time inference for AI models with trillions of parameters depend on this configuration attaining up to 83.2 peak petaflops of processing capacity, according to the tech giant.
AWS also revealed Project Rainier, a partnership with artificial intelligence safety and research group Anthropic. The project consists of building an EC2 UltraCluster out of hundreds of thousands of Trainium2 processors. Offering almost five times the computing capability utilized to train Anthropic's present models, the cluster seeks to establish a strong foundation for training and deploying next-generation AI models. Using this architecture, anthropic intends to maximize its Claude AI models for Trainium2, thereby improving its cost economy and performance.
Expecting to be on sale in late 2025, AWS showed its next generation Trainium3 AI chips. Designed on a 3-nanometer manufacturing node, these chips should provide four times the performance of Trn2 UltraServers. This progress will help to improve real-time inference skills and promote the creation of ever sophisticated artificial intelligence models.
Already using AWS Trainium technology are many business leaders. Trainium2 will be included into Databricks' Mosaic AI platform so that users may scale model training at lower prices. AWS Trainium has been included into Hugging Face, an open artificial intelligence development platform, therefore enabling developers to create and implement AI models quickly. Projecting cost reductions of up to 40% compared to current solutions, Poolside, a firm focusing on generative AI tools for developers, plans to teach its future models on Trn2 UltraServers, Amazon said.
Currently housed in the US East (Ohio) AWS Region, AWS Trainium 2 instances are scheduled to be shortly accessible in other regions. Early access to consumers hoping to grow their AI workloads comes via Trn2 UltraServers, which are under preview. The AWS News Blog has more details about Trainium 2 and other developments.
