NVidia's advantage is the tensor units. Not just software but also hardware.

MI100 finally brings hardware accelerated matrix-multiplication to the table (bfloat16, FP16, and even FP32 matrix multiplication).

Now that MI100 has matrix-multiplication units, now comes the part where that is distributed to the software. (Hooking up TensorFlow to use them).

--------

Without Matrix Multiplication units, there was really no point even writing software (aside from maybe preparation of the eventual rise of AMD's matrix-multiplication routines).

But 6800 XT won't have matrix-multiplication units. Only the CDNA architecture has those. So I'm not entirely sure if 6800 XT would even be an adequate competitor to deep learning applications, even if ROCm supported it entirely.

It's got nothing to do with hardware, so far NVIDIA has been the only vendor to take deep learning seriously. AMD GPUs have at times been the best on paper for DL but always lacked real software support. A low cost 16GB card could be very useful but their strategy doesn't appear to have changed so today NVIDIA is still the only option.

Isn't that what Google's TPU equipment is for?

And get sued by Google? You aren't allowed to use them to train anything that could compete with Google's business or products as part of the usage agreement.

Citation? I did a little digging of my own, because I found this claim to be uncharacteristic of any company I've ever worked for. I did find some restrictive clauses, but I don't think they are as broad as you are suggesting.

https://cloud.google.com/terms/service-terms Service Specific Terms -> AI and Machine Learning.

You can't use Google's fully-trained models as a reference in order to train your own models. Lower-level uses of TPUs to train your own models from your own training data are unrelated to those restrictions.

Certainly nothing in the Acceptable Use Policy looks anti-competitive to me.

Obvious disclaimer is obvious: IANAL.

So is the only reason to buy TPUs to learn about them and go work for Google or something? Why sell them if you can't use them for anything (as Google considers anything computing to be under their domain).

> Why sell them

They don't. GOOG rents them.

Sorry, when I said sell I was thinking of the Coral Edge TPU board.