[ResNet50/TensorFlow] Inconsistent benchmark result on two citations?

[372046933]

Related to Model/Framework(s)
The training script is from https://github.com/NVIDIA/DeepLearningExamples/blob/master/TensorFlow/Classification/RN50v1.5/main.py

Describe the bug
on https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html#stories_tensorflow Mixed precision of different models are given as follows.

Model	Speedup
BERT Q&A	3.3X speedup
GNMT	1.7X speedup
NCF	2.6X speedup
ResNet-50-v1.5	3.3X speedup
SSD-RN50-FPN-640	2.5X speedup

While training performance results on this repo is different:
https://github.com/NVIDIA/DeepLearningExamples/tree/master/TensorFlow/Classification/RN50v1.5#training-performance-benchmark

number of GPUs	mixed precision img/s	FP32 img/s	mixed precision speedup	mixed precision weak scaling	FP32 weak scaling
1	825.2	364.9	2.20	1.00	1.00
4	3197.4	1419.4	2.25	3.96	3.89
8	6209.9	2778.5	2.24	7.74	7.61

We can see that the mixed precision is 2.2 on this repository, but 3.3 on Nvidia's another site.

To Reproduce
Just read the documentation is enough

Expected behavior
Give consistent benchmark result.

Environment
Please provide at least:

• Container version (e.g. nvcr.io/nvidia/tensorflow:19.12-tf1-py3):
• GPUs in the system: (e.g. 8x Tesla V100-PCIE-16GB):
• CUDA driver version (e.g. 418.87):