Announcing Amazon EC2 Trn3 UltraServers for faster, lower-cost generative AI training

This article announces the general availability of Amazon EC2 Trn3 UltraServers powered by AWS Trainium3, a fourth-generation 3nm AI chip designed for cost-effective generative AI training.

• Trainium3 delivers 2.52 petaflops of FP8 compute with 144 GB HBM3e memory
• 1.5x more memory capacity and 1.7x higher bandwidth than Trainium2
• Scales up to 144 chips per UltraServer (362 FP8 PFLOPs total)
• 4.4x higher performance and 4x better performance/watt versus Trn2
• NeuronSwitch-v1 fabric doubles interchip interconnect bandwidth
• Native PyTorch integration enables training without code changes
• 3x faster performance on Amazon Bedrock with 5x higher output tokens per megawatt

Trn3 UltraServers provide best-in-class price-performance for training frontier-scale models including reinforcement learning, MoE, and reasoning applications.