Existing Likely

Microsoft GPT-4 cluster

Microsoft,OpenAI • United States of America • April 30, 2022
Total GPUs
25.0K
H100 Equiv
7.9K
Power
21.3 MW
FP16
7.8K PF

System Equivalents

This datacenter's 25.0K GPUs equal:

NVIDIA DGX A100
8 GPUs/system
3.1K
NVIDIA HGX A100 (8-GPU)
8 GPUs/system
3.1K

Specifications

Primary GPUNVIDIA A100
GPU Count25.0K
LocationDes Moines, Iowa
SectorPrivate

Additional Information

OpenAI’s training FLOPS for GPT-4 is ~2.15e25, on ~25,000 A100s for 90 to 100 days at about 32% to 36% MFU....

Training finished August 2022. Training lasted ~3 months, meaning that it was operational late May at the latest