Better explanation of raid 5:

You have 5GB of data and 5 disks. You split your data into 4 parts and split one on each disk. Then disk 5 remembers if there is an odd or even number of 1s on the other disks. So whichever disk fails you can count if it was odd or even. So you loose 1 disk but keep full capacity of the other disks. No doubling like suggested before

As someone who does AI for a living: GPU+docker is easy and reliable. Especially if you compare it to VMs.

But not the fancy machine learning acceleration