AWS Trainium2 reaches general availability, Trainium3 announced

2024-12-03 · Publisher : AWS · Domain : Cloud & DevOps

At re:Invent 2024, AWS declares Trn2 instances generally available with 30 to 40% better price-performance than its P5e and P5en GPU instances. It also announces Trainium3 and Project Rainier with Anthropic.

### What changes On 3 December 2024, at re:Invent, AWS announces general availability of Amazon EC2 Trn2 instances, built on its Trainium2 AI chip. AWS also presents Trn2 UltraServers (in preview), announces the next Trainium3 chip and unveils Project Rainier, a very large cluster built with Anthropic. AWS presents Trainium2 as built for both training and inference of the largest generative AI workloads. ### Key facts - Trn2 instance: 16 Trainium2 chips and 20.8 peak petaflops of compute; 30 to 40% better price-performance than the GPU-based EC2 P5e and P5en instances, according to AWS. - Trn2 UltraServer: 64 Trainium2 chips connected by NeuronLink, up to 83.2 peak petaflops of compute; in preview at announcement. - Trainium3: AWS's first chip on a 3 nm process, expected late 2025, with an announced performance gain of 4x over Trn2 UltraServers. - Project Rainier: an EC2 UltraCluster with hundreds of thousands of Trainium2 chips, for Anthropic's use. - Customers and partners quoted: Anthropic, Databricks (targeting 30% lower total cost), poolside (40% savings expected against P5 instances), Hugging Face. - Google is cited for supporting the integration of JAX with AWS's software stack. ### Why it matters AWS offers an alternative to NVIDIA GPUs for training and inference, at a lower announced price. For customers, the choice depends on software porting: the Neuron SDK is required, which takes engineering effort. The performance figures remain the vendor's own. ### Takeaway Hyperscalers turn their in-house chips into a real option against NVIDIA; Anthropic is the first major customer on the AWS side.

Tags : AWS, Puces, Hyperscaler, Anthropic, Conférence