Amazon Web Services Announces General Availability of AWS Trainium2 Instances

The new Trn2 instances are designed for optimal performance, boasting a 30-40% better price-performance ratio compared to the existing GPU-based EC2 P5e and P5en instances. Each Trn2 instance is equipped with 16 Trainium2 chips, providing a peak compute performance of 20.8 petaflops. This makes them particularly suited for training and deploying large language models (LLMs) that encompass billions of parameters.

AWS has also unveiled its Trn2 UltraServers, a completely new EC2 offering that features 64 interconnected Trainium2 chips. These UltraServers utilize the ultra-fast NeuronLink interconnect technology, allowing them to achieve a staggering 83.2 peak petaflops of compute. This innovative architecture enables customers to train and deploy the world’s largest AI models more efficiently, reducing both training times and costs.

In an ambitious collaboration with AI safety and research company Anthropic, AWS is developing an EC2 UltraCluster of Trn2 UltraServers dubbed Project Rainier. This project will consist of hundreds of thousands of Trainium2 chips and is expected to deliver over five times the number of exaflops utilized in Anthropic's current model training. When completed, it will potentially be the largest AI compute cluster ever built, enabling Anthropic to create and deploy future models at an unprecedented scale.

In addition to the Trainium2 announcement, AWS also unveiled its next-generation AI chip, the Trainium3. Expected to harness a 3-nanometer process node, Trainium3 is designed to meet the high-performance demands of next-generation generative AI workloads. The UltraServers powered by Trainium3 are projected to be four times more efficient than their Trn2 counterparts, allowing for faster model iteration and superior real-time performance.

AWS emphasized that the Trainium2 instances are purpose-built to support the most advanced generative AI workloads. David Brown, vice president of Compute and Networking at AWS, noted that as AI models grow in complexity, the need for innovative solutions to manage training and inference becomes crucial. The Trn2 UltraServers are expected to significantly enhance organizations' capabilities to train and deploy large-scale models quickly and cost-effectively.

Several leading companies in the AI space have already begun to harness the power of Trainium2. For instance, Databricks’ Mosaic AI is set to leverage Trainium2, predicting a reduction in total cost of ownership by up to 30%. Hugging Face, a dominant platform for AI developers, will also integrate Trainium's capabilities to enhance model performance and speed.

Trn2 instances are now generally available in the US East (Ohio) AWS Region, with plans for additional regions to follow soon. The Trn2 UltraServers are currently in preview, allowing select customers to experiment with the technology before broader deployment.

The launch of AWS Trainium2 and the associated UltraServers marks a bold step forward in the capabilities offered by AWS in the AI and machine learning arenas. By providing enhanced performance, scalability, and cost efficiency, AWS is positioning itself as a leader in facilitating the next generation of AI innovations. As the demand for sophisticated AI models continues to grow, AWS's developments are likely to play a pivotal role in shaping the future of AI technology.

For more information on this announcement, visit the AWS News Blog, the AWS Trainium page, and the AWS re:Invent page.