One would think that this announcement from Nvidia would speed up this development. Doesn't this new policy render tensorflow + Nvidia pretty much useless ?

No. Tesla and Quadro are definitely exempt from this potential update to the EULA, and you can have a lot of GeForce cards outside of a datacenter.

Then this is again the same kind of money grab as they have done with active stereo rendering - it used to work even on consumer-level hardware.

But then it got popular outside of research labs too, so Nvidia has decided to milk the cow and restrict this feature only to their Quadro line-up, with ATI following suit. The result has been that they have pretty much killed the market - apart from those research labs nobody is going to buy a $3-4k Quadro card that has otherwise the same performance as a $300 GeForce only for the stereo support.

I think this sort of artificial crippling/restricting of usage to force industrial customers to use the more expensive hw they don't need otherwise will only lead to a proliferation of task-specific ASICs and Nvidia will hurt only itself with it. The reason why people use Nvidia GPUs for parallel computing is cost and ubiquitous availability, not because there aren't other options. This move will only accelerate their development - for which there were no reasons until now.

A HIP version is available. https://github.com/ROCmSoftwarePlatform/hiptensorflow

Hip caffe is also quite well underway.

Why is it on AMD to support TensorFlow? Pretty sure it is on the machine learning people to use the standard APIs like OpenCL or Vulkan etc instead of CUDA...

The’re not running a non-profit. It’s on them to respond to the market.

Goes the other way around too. This whole debacle demonstrates how dumb it is to base your product on a proprietary API with just one vendor.

Here you go: https://github.com/tensorflow/tensorflow/issues/22

Companies doing ML do not and should not base their decisions on some hypothetical considerations about long-term effects on market health, or just present/future hardware cost. They work with what's available. It is market's job to correct. And it surely will now, as current situation becomes more of a problem.

Nobody wanted to do this, but the lousy state of OpenCL compared to CUDA left them no choice. It really is so far ahead that OpenCL was never an option. It’s on AMD to fix that, since NVIDIA certainly won’t.

No, because CUDA has first class C++ support, while Khronos only added C++ support when it became visible no one wants to keep using C, when better options are available in CUDA via PTX.

As with many things: it depends on who wants it more. In this case, I think AMD probably cares more.