Existing Confirmed

Paper on GLaM

Google • United States of America • December 13, 2021
Total GPUs
1.0K
H100 Equiv
142
Power
0.4 MW
FP16
282 PF

Specifications

Primary GPUGoogle TPU v4
GPU Count1.0K
SectorPrivate

Additional Information

"Moreover, to reach similar (and slightly exceeded) scores as GPT-3, we train using 1,024 TPU-v4 chips for 574 hours (with 280B tokens)."