There's a bunch of startups doing accelerators for ML/DL. Everybody trying a different take. Hardware is hard. It's expensive, difficult to develop and even more difficult for customers to try it (you can only give so many free samples for evaluation). Which begs the question: are they for real and want to look NVIDIA and Intel in the eye, or are they betting on an acquisition to survive in the long term?
From day 1 their goal is to make a big enough splash so that they get acquired by one of the big fishes. So far, there have been 2 instances of this.. both involving Intel being suckered into paying a ton of money for aqui-hiring these people: nervana for 350M in 2016, and Habana for few B last year
Yeah, but even Intel is not pushing them too: nervana products dissapeared and Habana's is unclear where they fit.
Yeah.. intel doesn't have a good home grown solution for ML acceleration and they can be easily baited into acquiring one of these "disruptive" startups. Afaik, there hasn't been any other major ML hw acquisition by other, more competant companies: amazon, facebook, google, apple.
Yeah its an inteesting domain no doubt. But not a lot of companies do DL, they do analytics but DL applications is pretty niche so I wonder why would a company like that not use specialised TPUs from Google/AWS instead of going to a startup.
The thinking is so Google doesn't gobble up more of their data. It is a common reply on why not Google.
GCP which has the TPU product doesnt gobble data. Thats enterprise data Google doesnt do data mining there.