Existing Confirmed

Paper on ViT

Google • June 8, 2021
Total GPUs
2.0K
H100 Equiv
127
Power
1.0 MW
FP16
252 PF

Specifications

Primary GPU Google TPU v3
GPU Count 2.0K
Sector Private

Additional Information

"For models ViT-g and ViT-G, to speed up the training, we scale up the batch size at most to 32 768 and distribute the training to 2048 TPUv3 chips."