Amazon EC2 instances powered by Trainium2 chips are now available

Amazon Web Services (AWS) has announced the launch of Amazon Elastic Compute Cloud (Amazon EC2) Trn2 compute instances, equipped with Trainium2 chips and designed specifically for generative Artificial Intelligence (AI) applications.

The company introduced Trainium2 at the previous edition of AWS re:Invent, where it announced that it was continuing to work to provide customers with more options to run any application or workload with both this chip and Graviton4.

AWS CEO Matt Garman announced at the 2024 annual conference, which began this Monday, the general availability of instances of its on-demand scalable computing capacity Amazon EC2 powered by AWS Trainium2.

In this sense, he recalled that the new Amazon EC2 Trn2 instances are specially designed for high-performance deep learning training of AI models, including large language models (LLM) and latent diffusion.

He also noted that Trn2 is the highest-performing Amazon EC2 instance for deep learning and generative AI, offering 30 to 40 percent better price performance than the current generation of graphics-based processing. on discs.

That’s because a single instance of Trn2 combines 16 new Trainium2 chips interconnected with NeuronLink’s ultra-fast bandwidth, as well as a low-latency chip-to-chip interconnect, with 20.8 petaflops of processing maximum.

AWS has also shown that for larger models that require more computing, it has worked on Trn2 UltraServers, virtual servers that allow software developers to scale processing capacity for any workload.

With these new systems, customers can scale their workloads to improve real-time inference performance for billion-parameter models in production. In this sense, the firm has indicated that it is building an EC2 UltraCluster of Trn2 Ultraservers together with Anthropic.

This project is called Project Rainier, which will extend distributed model training through Trainium2 chips connected to EFA networks, that is, network interfaces for Amazon EC2 instances that allow customers to run applications that require high levels of communication between nodes at scale. AWS.

Trn2 compute instances are already available in the AWS region in a part of the United States – specifically Ohio – and will soon be extended to other regions. Trn2 UltraServers, for their part, are in a preliminary version for testing.

TRAINIUM3

The company also announced at the meeting Trainium3, its next-generation AI chip, which will allow customers to create various larger models faster, as well as provide superior real-time performance in their implementation.

These will be the first AWS chips manufactured with a 3 nanometer (3nm) process node, setting a new standard for performance, power efficiency and density.

Thus, cloud servers used to train AI programs (UltraServers) with Trainium3 technology are expected to have four times the performance of Trn2 UltraServers. Thanks to this, customers will be able to iterate even faster when creating models and provide superior real-time performance in their deployments.

Amazon Web Services has finally announced that the first Trainium3 chips are expected to be available late next year.

Check Also

the element that cannot be missing in your wardrobe

Sports shoes are not only a practical accessory, but also an essential base in any …

Leave a Reply

Your email address will not be published. Required fields are marked *