Existing Likely

Paper on Switch transformer

Google • January 11, 2021
Total GPUs
1.0K
H100 Equiv
63
Power
0.5 MW
FP16
126 PF

Specifications

Primary GPU Google TPU v3
GPU Count 1.0K
Sector Private

Additional Information

The TPU count is from the expert parallelism (1 expert per core) with 2048 experts and 2 tensor cores per TPU v3.