云算力Cloud compute
AWS Trainium2(Trn2)AWS Trainium2 (Trn2)
AWS 自研训练加速器,第二代;单实例最多 16 颗,UltraServer 最多 64 颗,用 NeuronLink 互联。软件栈是 AWS Neuron,不是 CUDA,迁移成本要单列。页面没有给 Trainium2 单芯片 HBM 容量,本站留空不猜;同页的 Trainium3 是下一代(144 GB HBM3e、4.9 TB/s),不要把两代数字混在一张卡里。AWS’s second-generation in-house training accelerator: up to 16 chips per instance and 64 per UltraServer, linked by NeuronLink. The software stack is AWS Neuron, not CUDA, so migration cost is a separate line item. The page does not list per-chip HBM for Trainium2, so this card leaves it blank rather than guessing; Trainium3 on the same page is a different generation (144 GB HBM3e, 4.9 TB/s).
关键参数Key specs
- 芯片数 / chips
- up to 16 per Trn2 instance; up to 64 per Trn2 UltraServer
- 互联 / interconnect
- NeuronLink
- 相对性能 / vs Trainium1
- 4x
- 性价比 / price-performance
- 30–40% better than P5e/P5en (AWS claim)
- 软件栈 / software
- AWS Neuron SDK
- 单芯片 HBM / HBM per chip
- not stated on page
官方原文怎么说What the source says
AWS Trainium2 芯片的性能是第一代 Trainium 的 4 倍。
Trn2 实例采用多达 16 个 Trainium2 芯片,Trn2 UltraServer 采用多达 64 个 Trainium2 芯片,这些芯片使用我们专有的芯片间互连技术 NeuronLink 进行互连。
与基于 GPU 的 EC2 P5e 和 P5en 实例相比,它们的性价比提高了 30% 到 40%。
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-17
- 许可 / 类型Licence / kind
- official-spec-page
- 许可说明Licence evidence
- https://aws.amazon.com/machine-learning/trainium/
- 类目Category
- 云算力Cloud compute