AWS Trainium2
芯片
AI加速器
发货中
The chip under Project Rainier — Anthropic's half-million-chip training fleet.
您正在查看一条芯片记录：该处理器是什么、其关键规格、谁设计它、谁制造它、谁供应其零部件，以及它被用在哪里。
Trainium2 is AWS's frontier-scale training part: 96 GB of HBM3 per chip, 1.3 dense FP8 PFLOPS, and 64-chip 'UltraServers' wired with NeuronLink. Its defining deployment is Project Rainier — the multi-site cluster AWS built with and for Anthropic, described at launch as approaching 500,000 chips and reported to be expanding toward a million by end of 2025. It is the largest non-NVIDIA AI build-out publicly known.
规格与关键参数
每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。
Overview
Designer
AWS (Annapurna Labs)
GA
December 2024
Status
Shipping
内存
Memory
96 GB HBM3Vendor spec
Memory bandwidth
2.9 TB/sVendor spec
算力
FP8 (dense)
1.3 PFLOPSVendor spec
Sparse FP8
5.2 PFLOPS (16:4 sparsity)Vendor spec
Physical
UltraServer
64 chips, 83.2 FP8 PFLOPS dense, 6 TB HBMVendor spec
Interconnect
NeuronLink; EFA scale-out networkingVendor spec
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。 +3 更多分组见下方。
由……制造
内存技术
内部包含
后继产品
前代产品
来源
AWS Trainium2 GA announcement ↗
记录审阅截至 2026-01-15 — 查看 方法论.