Trainium
#1amazon · seit Dezember 2025 (Trainium3) · 16× · zuletzt 30. Juni 2026
100
Momentum
Trainium is AWS's purpose-built AI accelerator chip for training and inference of large language models. The current generation (Trainium3), launched in December 2025 on TSMC's 3nm process, delivers 2.52 PFLOPs of FP8 compute per chip with 144 GB HBM3e memory. AWS positions Trainium as a cost-effective alternative to NVIDIA chips, offering 30-50% savings in total cost of ownership for customers.
Historique du momentum
04.04.03.07.
Fonctionnalités
| Manufacturing Process (nm) | 3 nm (Trainium3, first 3nm AWS chip); Trainium2: 5 nm; Trainium1: 7 nm |
| License | Neuron Kernel Interface (NKI) compiler under Apache 2.0 open source; chip/hardware itself proprietary, available only as an AWS cloud service |
| Platform | Amazon EC2 (Trn1/Trn2/Trn3 instances & UltraServer), programmable via AWS Neuron SDK, compatible with PyTorch, JAX, Hugging Face, vLLM |
| Price | Trn3: approx. $1.80/chip-hour (third-party source); Trn1.32xlarge from $21.50/h on-demand |
| Compute Performance (FLOPS/TOPS) | Trainium3: 2.52 PFLOPS FP8 per chip; Trn3 UltraServer (144 chips): up to 362 PFLOPS FP8/MXFP8 |
| Release Date | Trainium3 / Trn3 UltraServer GA: December 2, 2025 (AWS re:Invent 2025) |
| Memory | 144 GB HBM3e per chip, 4.9 TB/s bandwidth; UltraServer up to 20.7 TB HBM3e, 706 TB/s aggregate bandwidth |
| Availability | Trainium3/Trn3 UltraServers generally available (GA) since December 2, 2025 via AWS EC2 |
