Very slow execution time with NC6_Standart

Julien 1

Hello,
In Azure ML, I created an NC6_Standart instance in order to have GPU. In order to familiarize myself with the cloud, I decide to run this (which is actually in a microsoft tutorial) https://github.com/MicrosoftDocs/tensorflowfundamentals/blob/main/intro-keras/kintro.py on my pc it takes about 2~3 minutes to do 100 epochs (with 2ms/step), from the documentation sandbox about 3 minutes also with 2ms/step but from the NC6 instance it takes 8 minutes BUT with 2ms/step. In fact, it is the time between each epoch that is long. Would you have a solution?

romungi-MSFT 43,696 Reputation points Microsoft Employee

2022-09-01T10:59:08.82+00:00

@Julien All the compute used by Azure ML experiments are offered as is with the specification as available under Azure compute VM types.
I think by default the GPU capabilities of the VM are not used by the default image of Azure ML, you might want to look at using a GPU image and try to setup the experiment and submit it to Azure ML.

https://learn.microsoft.com/en-us/azure/machine-learning/how-to-train-keras#define-your-environment

Share via

Very slow execution time with NC6_Standart