4 comments

[ 0.22 ms ] story [ 23.3 ms ] thread
The real question is: does Google Cloud charge more or less per GPU cycle than one can make mining cryptocurrency with it? If it's less, I expect these to be at load momentarily.
They charge more. It's significantly cheaper to build your own infrastructure if you don't need the on demand availability, reliability and minimal maintenance the cloud offers.
Can someone explain me why one would need a container orchestration system to train ml models? Why wouldn't one just stick with containers?
For the same reason you would use a container orchestration system for running anything: easier deployments, scaling, reliability, etc.