
Azure Innovations with Mark Russinovich
RunAs Radio
00:00
How to Train Machine Learning Models for the Cloud
The training requires a type of data parallelism. You take many instances of the model you're going to train, and then you feed them all chunks of data. And that's why you want these large clusters because the more GPUs you can have running in this data parallel mode, the faster your training will be. At the inference side or the serving side, once you've got the trained model, of course, you can just deploy it once and use it. Right. We know how much bigger GBD is for us. So extrapolate at your peril.
Transcript
Play full episode