Really great breakdown! Another important thing to think about when building out a production or even dev environment are the other associated costs (i.e. bandwidth, storage, egress, etc). We think that these costs can be pretty obscene in the public cloud and we try to optimize as much as possible to keep the total cost down.
Disclosure: I'm one of the founders of Paperspace (https://www.paperspace.com) and we spend a lot of time thinking about GPU compute and pricing. Happy to answer any questions here
- How much time (=cost) do you need to start running (enterprise)
- Time you need to access the cloud, starting up and closing vms, and getting shared GPU resources.
- The cost of lock-in. Ecosystems are great; but sometimes you need the basics working.
Some providers are much better than others at this.
“IBM Softlayer and LeaderGPU appear expensive, mainly due to under-utilisation of their multi-GPU instances. The benchmark was carried out using the Keras framework whose multi-GPU implementation was surprisingly inefficient, at times performing worse than a single GPU run on the same machine.“ - this is unacceptable in a benchmark like this. There is an entire software stack that is influencing performance and much of the hardware is dissimilar.
Disclosure: I'm one of the founders of Paperspace (https://www.paperspace.com) and we spend a lot of time thinking about GPU compute and pricing. Happy to answer any questions here