List your company

Private Transformer Inference Efficiency on Transformer Block BERT-base settings

3.05Communication Cost (MB)Nimbus
-24.406160.922346.25531.578Nov 24, 2024

Evaluation results

MethodLinks
NimbusLayer=Linearo2024.113.051
NimbusLayer=Linearh22024.113.051
BumbleBeeLayer=Linearo2024.116.712
NimbusLayer=Linearqkv2024.1110.351
BumbleBeeLayer=Linearqkv2024.1114.472
NimbusLayer=Linearh12024.1115.521
BumbleBeeLayer=Linearh22024.1115.712
BumbleBeeLayer=Linearh12024.1118.352
IronLayer=Linearo2024.1140.22
NimbusLayer=GELU2024.1153.2224
IronLayer=Linearqkv2024.1174.642
IronLayer=Linearh22024.1178.372
IronLayer=Linearh12024.1184.462
NimbusLayer=Softmax2024.11115.3560
BumbleBeeLayer=GELU2024.11185.1346
IronLayer=GELU2024.11283.8965
BumbleBeeLayer=Softmax2024.11354.2670
IronLayer=Softmax2024.11689.45110