Table 4.

Computational efficiency metrics for BERT QA training.

Model	FLOPs	Loss	Runtime (s)	SPS	STS
BERT uncased [44]	$3.54\times {10}^{10}$	$0.000100$	77,846.59	$0.18$	$0.012$
RoBERTa [45]	$3.53\times {10}^{10}$	$0.000000$	77,985.63	$0.19$	$0.011$
DistilBERT [46]	$1.76\times {10}^{10}$	$0.004300$	41,347.49	$0.35$	$0.022$