Linus has been wrong before, but here he is clearly and unambiguously wrong. You don't - you NEVER - develop to the details of an architecture. The only exception is when you're doing OS development. Since things that people want to run in the "cloud" are exceedingly rarely OS development, the idea that you'd care about the architecture is quite silly.
We have vendors ardently trying to get us to commit to their specific environments, like NVIDIA, but they've been dropping bricks on their feet by not actively encouraging software developers to not target specific versions of CUDA. Anyone running TensorFlow knows how much this sucks. After a while, people are going to get fed up with the problems that come from a poor yet popular implementation and are going to start preferring less popular, yet less proprietary, more deployable solutions.
The idea that Intel and AMD can continue to keep the x86 platform performance advantageous forever is pretty ridiculous. We are already seeing Arm CPUs which are significantly better than x86 in performance per watt. How long will it be until we have high end consumer Arm and low end server Arm that completely overlap x86? One year? Two? Three? Possibly four?
x86 isn't over - that's not what's happening here. But it certainly hasn't "won" forever.
> You don't - you NEVER - develop to the details of an architecture.
You don't try to develop to the details of an architecture. You even try not to. Then you try to run it on a different architecture, and you find (some of) the places you developed to your specific architecture even though you didn't mean to.
And that's why Linus is right. It's easy to be architecture-specific by accident. It's really hard to not. And it's going to take time and effort to go to a different architecture. In the real world, few people want to waste their time doing that.
We have vendors ardently trying to get us to commit to their specific environments, like NVIDIA, but they've been dropping bricks on their feet by not actively encouraging software developers to not target specific versions of CUDA. Anyone running TensorFlow knows how much this sucks. After a while, people are going to get fed up with the problems that come from a poor yet popular implementation and are going to start preferring less popular, yet less proprietary, more deployable solutions.
The idea that Intel and AMD can continue to keep the x86 platform performance advantageous forever is pretty ridiculous. We are already seeing Arm CPUs which are significantly better than x86 in performance per watt. How long will it be until we have high end consumer Arm and low end server Arm that completely overlap x86? One year? Two? Three? Possibly four?
x86 isn't over - that's not what's happening here. But it certainly hasn't "won" forever.