Don't prematurely optimize. ML these days has all sorts of scale and scope. For instance, what used to take a few days baking on a GPU server can now be run in

Don't prematurely optimize. ML these days has all sorts of scale and scope. For instance, what used to take a few days baking on a GPU server can now be run in a few lines of codes on a CPU. The ecosystem moves very quickly, so it's better to be adaptable than to consign yourself to a specific paradigm of computation. That being said: Yeah, a lot of stuff is in CUDA or nvidia world. AMD and OpenCL support is usually late and usually lackluster. These days you can do a lot free to get started. Get started with Kaggle kernels and Colab for free. They give you access to nice GPUs. Learn a little bit of how to do things on AWS/GCP, because almost always people are just trying to burn their free credits as part of their investment package or something. If you're utterly concerned with performance, most workstations I know of usually run Ubuntu. Performance is far superior. Owning or building a GPU workstation is still remarkably cheaper than renting to some degree. A key idea I'd like to highlight here is that with ML/AI, regardless if your data is small or your batch processes "doable", you should be automating your pipeline so you don't have to worry about it. In the same sense that a GROUP-BY SQL statement is also machine learning speaking roughly. Being able to get your data into workable shape is more important than having nice hardware. Get the data to work for you first before you think about performance or speed.