Projects like training an ML model break down into compute and storage, the former of which is charged hourly and the latter is charged per gigabyte-month. Because of the variability in ML model complexity it’s hard for us to say in the general case how much it would cost to retrain something like a GAN, but if you can identify:
- roughly how long (in hours) it takes you to retrain a model on local hardware
- how much training data you need
you can then correlate these numbers to the pricing pages provided by the cloud vendors:
- Compute
- AWS virtual machine descriptions: Instance Type Explorer - Amazon Web Services
- AWS virtual machine pricing: EC2 On-Demand Instance Pricing – Amazon Web Services
- Microsoft Azure virtual machine pricing: Pricing - Linux Virtual Machines | Microsoft Azure
- Azure virtual machine selection wizard: Microsoft Azure VM Selector
- Storage
- AWS file storage pricing: Serverless File System | Amazon Elastic File System Pricing | AWS
- Azure file storage pricing: https://azure.microsoft.com/en-us/pricing/details/storage/files/
So purely for example – if it takes you 48 hours to train a model on a machine with 4 CPU cores, 16 GB of RAM and a good GPU, and you need 80 GB worth of training data, we can look up the cost of an equivalent virtual machine and that much data storage. On AWS, we’d look through their VM descriptions for something that references machine learning, and look for things on the lower end of the price range. Maybe we’d settle on a VM of type “g4dn.xlarge”, which costs $.526 to rent per hour. Storage will cost us $.43 per GB per month. So now we have something like:
($.526/vm-hour * 48 hours) + ($.043/gb-month * 80 GB) ~= $26 compute + $4 storage ~= $30
Keep in mind the storage cost is per month, so once your data is on the cloud the only cost you’re racking up for the rest of the month is for training runs.