Misclassification error of example mnist _mlp.py with GPU backend is abnormally high

[StarGazerM]

Environment:

• python -- v3.5.6
• neon -- 2.6.0
• cuda -- 10.0.130
• os -- manjaro linux 18.0.2
• gpu -- GTX 1050 mobile

Problem
After installation, we I try to test with mnist_mlp.py in examlpe. The misclassification becomes 90.2%,
while no error occur.
example run normally with other backend.