GPU compute cluster faster on Windows 10 than Ubuntu 18.04

Question

Info: My windows and Ubuntu system is the same. I just swap out my boot drive using a removable drive bay.

*Not using Docker

CPU: i9-7920X (128 GB DDR4)

GPUs: 2 Titan Vs and RTX 2070 (Display)

Windows 10: Nvidia Driver 416.81 in TCC cluster and Cuda 10

Ubuntu Desktop 18.04: Nvidia Driver 410.72 (not sure if need to cluster on Linux) and Cuda 10

Test Software: Running TensorFlow cifar10 esitimator - same training data - (link at bottom)

python cifar10_main.py --data-dir=/tmp --job-dir=/tmp/cifar10 --num-gpus=2 --train-steps=10000

Windows completes in 420 seconds (7.0 minutes) Ubuntu completes in 614 seconds (10.2 minutes)

https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10_estimator

Why is Ubuntu so much slower? Do the GPUs need to be clustered on Linux or are the different drivers (Ubuntu vs. Windows) the issue?

Thanks.