But that's not the world we live in. Nvidia's pitch is "We make the only hardware that runs the framework used by almost all deep learning research and media creation software, and we're also the only folks operating at our level when it comes to high-end video game graphics, and if our high-end cards are too expensive for you, we have cheaper ones. And when our competitors start to think they can catch up, we'll drop RTX and Tensor Cores on their face."
AMD seems stuck on "We have great value middle-to-high-tier video game graphics cards" at best. I have no idea how they can get out from under that rock. They've been turning the CPU market upside down and smashing Intel's once proud "near monopoly" status. Nvidia seems like exactly the kind of prideful company that would be poised to fall, but I have no idea how AMD could make it happen.
Or that it was cheaper to get reliability at another layer of the stack (rather than hardware).
Lose-lose for customers.
Most people aren't spending over $500 on a GPU, so they get the volume sales. The better aftermarket cards have been selling out pretty consistently. And the longer term strategies are similar to how they approached Ryzen. So, I'd think that Navi can definitely succeed in that light.
The real lock in for NVidia, is all the existing CUDA development. Intel and AMD will have a lot of work, One API may help there, so long as Intel and AMD can work together, without Intel's often and weird lock in attempts.
It was always about software. Granted, CUDA is not the best and most elegant platform in the world... but AMD seems not to be able to reach even that level.
I'd still likely only go AMD GPUs in the near future, just because I don't have particularly demanding GPU requirements and they have better Linux drivers.
Anyone privy on what hold Apple back on supporting Nvidia? The only info I read was due to some earlier bad blood, but I wonder why hurt your users just because of some old scuffles...
They probably both each want too much money.
So no, it is not a replacement, again. Elegance matters.
For and Engineer or Computation Scientist, I think this is not a bad investment
But Yes existing CUDA libraries won't work.