Hacker News Reader

> I think attempts from e.g. Google to develop their own chips will fail

I'm curious what you think failure will look like. In a lot of ways, Google's TPU program has already been extremely successful.

Interesting. Do you think this is more interpolability between packages, i.e. Pytorch not being written or having extensive testing for TPUs, or most deep learning problems don’t fit neatly into the exact problem domain TPU is specialized for?

I've also been seeing lots of gpu databases kind of hitting the mainstream( Mapd, brytlyt , blazingdb, Kinetica ). Not sure how big of a market that is.