> One thing Nvidia are much better at is supporting the whole stack - while I doubt many people are actually running much ML on 8gb mid-low end cards, it seems

> One thing Nvidia are much better at is supporting the whole stack - while I doubt many people are actually running much ML on 8gb mid-low end cards, it seems a hook for sales if people "could" do that, and a starting point for personal experimentation. I know multiple people who buy low end Nvidia GPUs because they "might" experiment with CUDA some time in the future. but this is very important. Many develop locally before sending it to a cluster, especially if you only have limited credits available.