Existing Likely

Amazon Titan training cluster

Amazon • United States of America • January 15, 2023
Total GPUs
13.8K
H100 Equiv
4.3K
Power
11.4 MW
FP16
4.3K PF

System Equivalents

This datacenter's 13.8K GPUs equal:

NVIDIA DGX A100
8 GPUs/system
1.7K
NVIDIA HGX A100 (8-GPU)
8 GPUs/system
1.7K

Specifications

Primary GPUNVIDIA A100
GPU Count13.8K
SectorPrivate

Additional Information

Specifically, they trained a 200B dense model on 4T tokens of data across 13,760 NVIDIA A100 chips (using 1,720 P4d nodes)

Model was released to large clients April 2023, took 48 days to train, so this cluster was operational at least three months beforehand