Google plans to sell the TPU as a cloud service. It's also their secret weapon, so to speak. They don't need people to buy it in order for it to be an asset, th

Google plans to sell the TPU as a cloud service. It's also their secret weapon, so to speak. They don't need people to buy it in order for it to be an asset, they mostly want to give themselves an advantage. That is, all the people at DeepMind now have access to faster hardware with more memory. They can run bigger experiments than everyone else. As for memory hierarchy, in hardware, there is always something to gain by specialization. Something like a convolution in a CNN is going to have very predictable memory access patterns. It's possible to take advantage of that. Google claims to already have about 10x speedups over GPUs IIRC.