AWS · cloud-model
Verified 2w ago
Amazon's second-generation training/inference silicon: HBM3 per chip; UltraServers via NeuronLink scale to dozens of chips.
Trainium2 is AWS's second-generation in-house AI accelerator, succeeded by Trainium3 in December 2025. Each chip pairs eight third-gen NeuronCores with 96 GB of HBM3 at 2.9 TB/s. Trn2 instances host 16 chips for 20.8 FP8 PFLOPS; UltraServers connect 64 chips for 83.2 FP8 PFLOPS.
AWS in-house AI silicon. Available through EC2 Trn2 instances and UltraServers; no retail SKU.
AWS's current generation is Trainium3. Trn3 UltraServers became generally available on 2025-12-02, with 144 GB of HBM3e per chip at 4.9 TB/s and up to 144 chips (362 FP8 PFLOPS) per UltraServer. Trn2 instances remain on offer.