Infrastructure atlas
AWS/Neuron accelerator

Trainium2

AWS’s second-generation training accelerator, exposed through EC2 Trn2 and UltraServer systems with Neuron software.

ASICTrainiumtraininginferenceAWS Neuron
Trainium2 hardware

01 / Memory

96 GiB HBM

Per chip

02 / Bandwidth

2.9 TB/s

Per chip

03 / Instance

16 chips

Trn2.48xlarge

Precision & peak compute

PrecisionPublished peakBasis
FP8Up to 1.3 PFLOPSDense, per chip

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

AWS Trainium2 and Trn2

Technical specification

Chip & compute
Core design
8 NeuronCores per chip
Dense FP8
Up to 1.3 PFLOPS per chip
HBM
96 GiB · 2.9 TB/s
Software
AWS Neuron SDK
EC2 system
Trn2 instance
16 chips · 128 NeuronCores
Instance HBM
1.5 TiB · 46 TB/s
Chip interconnect
NeuronLink · 1 TB/s
UltraServer
4 Trn2 instances · 64 chips