AWS Unleashes Trainium3 Chip: 4.4x Faster AI Performance for EC2 UltraServers

AWS has announced the launch of its Amazon EC2 Trn3 UltraServers, powered by a new 3 nm AI chip, Trainium3, designed to tackle rising cost and compute bottlenecks in AI training and inference workloads. According to official metrics, Trainium3 delivers up to 4.4× the performance of the previous generation while improving energy efficiency by 40%, and AWS positions the platform at a more competitive price point for large‑scale AI model training.

The EC2 Trn3 UltraServer can host up to 144 Trainium3 chips per system, offering 362 FP8 PFLOPs of AI compute, and is available starting today. To address communication bottlenecks, AWS has introduced NeuronSwitch‑v1 alongside an enhanced Neuron Fabric interconnect, reducing chip‑to‑chip latency to below 10 microseconds. Through EC2 UltraClusters 3.0 design, customers can interconnect thousands of UltraServers, scaling up to ultra‑large clusters containing one million Trainium chips.