re:Invent 2025: AWS Trainium3 UltraServers reach general availability

2025-12-02 · Publisher : AWS · Domain : Cloud & DevOps

At re:Invent 2025, AWS declares its EC2 Trn3 UltraServers generally available, built on Trainium3, its first 3 nm AI chip. AWS claims up to 4.4x the performance of Trn2 UltraServers, with 144 chips per server.

### Background On 2 December 2025, during re:Invent, AWS announces general availability of Amazon EC2 Trn3 UltraServers, equipped with Trainium3, its fourth-generation AI chip and the first built on a 3 nm process. The announcement follows December 2024, when AWS had promised Trainium3 for late 2025. AWS also presents Graviton5, its general-purpose processor, in preview. ### Key facts - Per Trainium3 chip: 2.52 petaflops in FP8, 144 GB of HBM3e memory and 4.9 TB/s of memory bandwidth. - A Trn3 UltraServer assembles up to 144 chips (362 FP8 petaflops), with 20.7 TB of HBM3e and 706 TB/s of aggregate memory bandwidth; UltraClusters 3.0 can reach hundreds of thousands of chips. - According to AWS: up to 4.4x the performance, 3.9x the memory bandwidth and 4x the performance per watt of Trn2 UltraServers. - On Amazon Bedrock, Trainium3 already serves production workloads, with up to 3x the performance of Trainium2. - Graviton5 (preview): 192 cores per chip and up to 25% more performance than the previous generation, according to AWS. - Customers such as Anthropic, Ricoh and Metagenomi report up to 50% lower training and inference costs with Trainium. ### Why it matters AWS maintains an annual cadence of renewal for its AI chips and in-house processors. For teams that train or serve models, it makes a non-NVIDIA-GPU choice credible, at the cost of software porting through the Neuron SDK. The gains quoted are the vendor's own and depend on workloads. ### Takeaway Trainium3 arrives as planned: AWS keeps its AI silicon roadmap a year after announcing it.

Tags : AWS, Puces, Hyperscaler, IA, Conférence