At AWS re:Invent, Amazon Web Services, Inc. (AWS), an Amazon.com, Inc. company, today announced the general availability of AWS Trainium2-powered Amazon Elastic Compute Cloud (Amazon EC2) instances, introduced new Trn2 UltraServers, enabling customers to train and deploy today's latest AI models as well as future large language models (LLM) and foundation models (FM) with exceptional levels of performance and cost efficiency, and unveiled next-generation Trainium3 chips.

"Trainium2 is purpose built to support the largest, most cutting-edge generative AI workloads, for both training and inference, and to deliver the best price performance on AWS," said David Brown, vice president of Compute and Networking at AWS. "With models approaching trillions of parameters, we understand customers also need a novel approach to train and run these massive workloads. New Trn2 UltraServers offer the fastest training and inference performance on AWS and help organizations of all sizes to train and deploy the world's largest models faster and at a lower cost."

• Trn2 instances offer 30-40% better price performance than the current generation of GPU-based EC2 P5e and P5en instances and feature 16 Trainium2 chips to provide 20.8 peak petaflops of compute—ideal for training and deploying LLMs with billions of parameters.
• Amazon EC2 Trn2 UltraServers are a completely new EC2 offering featuring 64 interconnected Trainium2 chips, using ultra-fast NeuronLink interconnect, to scale up to 83.2 peak petaflops of compute—quadrupling the compute, memory, and networking of a single instance—which makes it possible to train and deploy the world's largest models.
• Together with Anthropic, AWS is building an EC2 UltraCluster of Trn2 UltraServers—named Project Rainier—containing hundreds of thousands of Trainium2 chips and more than 5x the number of exaflops used to train their current generation of leading AI models.
• AWS unveiled Trainium3, its next generation AI chip, which will allow customers to build bigger models faster and deliver superior real-time performance when deploying them.

"Trainium2 is purpose built to support the largest, most cutting-edge generative AI workloads, for both training and inference, and to deliver the best price performance on AWS," said David Brown, vice president of Compute and Networking at AWS. "With models approaching trillions of parameters, we understand customers also need a novel approach to train and run these massive workloads. New Trn2 UltraServers offer the fastest training and inference performance on AWS and help organizations of all sizes to train and deploy the world's largest models faster and at a lower cost."

As models grow in size, they are pushing the limits of compute and networking infrastructure as customers seek to reduce training times and inference latency—the time between when an AI system receives an input and generates the corresponding output. AWS already offers the broadest and deepest selection of accelerated EC2 instances for AI/ML, including those powered by GPUs and ML chips. But even with the fastest accelerated instances available today, customers want more performance and scale to train these increasingly sophisticated models faster, at a lower cost. As model complexity and data volumes grow, simply increasing cluster size fails to yield faster training time due to parallelization constraints. Simultaneously, the demands of real-time inference push single-instance architectures beyond their capabilities.

Trn2 is the highest performing Amazon EC2 instance for deep learning and generative AI
Trn2 offers 30-40% better price performance than the current generation of GPU-based EC2 instances. A single Trn2 instance combines 16 Trainium2 chips interconnected with ultra-fast NeuronLink high-bandwidth, low-latency chip-to-chip interconnect to provide 20.8 peak petaflops of compute, ideal for training and deploying models that are billions of parameters in size.

Trn2 UltraServers meet increasingly demanding AI compute needs of the world's largest models
For the largest models that require even more compute, Trn2 UltraServers allow customers to scale training beyond the limits of a single Trn2 instance, reducing training time, accelerating time to market, and enabling rapid iteration to improve model accuracy. Trn2 UltraServers are a completely new EC2 offering that use ultra-fast NeuronLink interconnect to connect four Trn2 servers together into one giant server. With new Trn2 UltraServers, customers can scale up their generative AI workloads across 64 Trainium2 chips. For inference workloads, customers can use Trn2 UltraServers to improve real-time inference performance for trillion-parameter models in production. Together with Anthropic, AWS is building an EC2 UltraCluster of Trn2 UltraServers, named Project Rainier, which will scale out distributed model training across hundreds of thousands of Trainium2 chips interconnected with third-generation, low-latency petabit scale EFA networking—more than 5x the number of exaflops that Anthropic used to train their current generation of leading AI models. When completed, it is expected to be the world's largest AI compute cluster reported to date available for Anthropic to build and deploy their future models on.

Anthropic is an AI safety and research company that creates reliable, interpretable, and steerable AI systems. Anthropic's flagship product is Claude, an LLM trusted by millions of users worldwide. As part of Anthropic's expanded collaboration with AWS, they've begun optimizing Claude models to run on Trainium2, Amazon's most advanced AI hardware to date. Anthropic will be using hundreds of thousands of Trainium2 chips—over five times the size of their previous cluster—to deliver exceptional performance for customers using Claude in Amazon Bedrock.

Databricks' Mosaic AI enables organizations to build and deploy quality agent systems. It is built natively on top of the data lakehouse, enabling customers to easily and securely customize their models with enterprise data and deliver more accurate and domain-specific outputs. Thanks to Trainium's high performance and cost-effectiveness, customers can scale model training on Mosaic AI at a low cost. Trainium2's availability will be a major benefit to Databricks and its customers as demand for Mosaic AI continues to scale across all customer segments and around the world. Databricks, one of the largest data and AI companies in the world, plans to use Trn2 to deliver better results and lower TCO by up to 30% for its customers.

Hugging Face is the leading open platform for AI builders, with more than 2 million models, datasets, and AI applications shared by a community of more than 5 million researchers, data scientists, machine learning engineers, and software developers. Hugging Face has collaborated with AWS over the last couple of years, making it easier for developers to experience the performance and cost benefits of AWS Inferentia and Trainium through the Optimum Neuron open-source library, integrated in Hugging Face Inference Endpoints and now optimized within the new HUGS self-deployment service, available on the AWS Marketplace. With the launch of Trainium2, Hugging Face users will have access to even higher performance to develop and deploy models faster.

poolside is set to build a world where AI will drive the majority of economically valuable work and scientific progress. poolside believes that software development will be the first major capability in neural networks that reaches human-level intelligence. To enable that, they're building FMs, an API, and an assistant to bring the power of generative AI to developers' hands. A key to enable this technology is the infrastructure they're using to build and run their products. With AWS Trainium2, poolside's customers will be able to scale their usage of poolside at a price performance ratio unlike other AI accelerators. In addition, poolside plans to train future models with Trainium2 UltraServers, with expected savings of 40% compared to EC2 P5 instances.

Trainium3 chips—designed for the high-performance needs of the next frontier of generative AI workloads
AWS unveiled Trainium3, its next-generation AI training chip. Trainium3 will be the first AWS chip made with a 3-nanometer process node, setting a new standard for performance, power efficiency, and density. Trainium3-powered UltraServers are expected to be 4x more performant than Trn2 UltraServers, allowing customers to iterate even faster when building models and deliver superior real-time performance when deploying them. The first Trainium3-based instances are expected to be available in late 2025.

Enabling customers to unlock the performance of Trainium2 with AWS Neuron software
The Neuron SDK includes compiler, runtime libraries, and tools to help developers optimize their models to run on Trainium. It provides developers with the ability to optimize models for optimal performance on Trainium chips. Neuron is natively integrated with popular frameworks like JAX and PyTorch so customers can continue using their existing code and workflows on Trainium with fewer code changes. Neuron also supports over 100,000 models on the Hugging Face model hub. With the Neuron Kernel Interface (NKI), developers get access to bare metal Trainium chips, enabling them to write compute kernels that maximize performance for demanding workloads.

Neuron software is designed to make it easy to use popular frameworks like JAX to train and deploy models on Trainium2 while minimizing code changes and tie-in to vendor-specific solutions. Google is supporting AWS's efforts to enable customers to use JAX for large-scale training and inference through its native OpenXLA integration, providing users an easy and portable coding path to get started with Trn2 instances quickly. With industry wide open-source collaboration and the availability of Trainium2, Google expects to see increased adoption of JAX across the ML community—a significant milestone for the entire ML ecosystem.

Trn2 instances are generally available today in the US East (Ohio) AWS Region, with availability in additional regions coming soon. Trn2 UltraServers are available in preview.