Tensorflow ssd_resnet50 runs faster on tensorflow-gpu than XLA enabled 19.07-py2 container !

Hi, have noticed that ssd_resnet50 (from Tensorflow model zoo) runs faster on vanilla tensorflow:devel-gpu containers than on 19.07-py2 containers by a significant amount. In fact the same code runs faster on 19.04-py2 containers. The following are the numbers (in seconds for single image) on T4.

Container                    Non-AMP                AMP

tensorflow-gpu:devel-gpu 0.0854758467197418          N/A
19.04-py2                0.10218918190002442         0.08253747978210449
19.07-py2                0.104998300004005           0.0920532138109207