BERT
Benchmarks
TaskDatasetSOTA resultTrendResults
BERT-Base (train)
3,707.01Memory (MB)
15
BERT Base (inference)
18.591Total Time (s)
14
BERT (val)
65.54Accuracy
14
BERT-Large
22.6Tuning Time (s)
9
BERT Base
23.3Tuning Time (s)
9
BERT-base
40.54Attention Layer Latency (s)
6
BERT large
0.002vNMSE
6
BERT
0.003Latency per G
6
BERT-Large NVIDIA V100 (train)
96Max Batch Size
6
BERT-Large NVIDIA 2080 Ti (train)
50Max Batch Size
6
BERT base
21.5Speedup
5
BERT-base
2.2Online Overhead (GB)
4
BERT Large (inference)
0.351GeLU Time (s)
4
BERT
53.29Accuracy
3
BERT base
589,824Nova
1
BERT MLM small (val)
6.941Validation Loss
1