"CUDA is still a giant moat for Nvidia"
"It's really difficult to explain to non-programmers -- or even programmers who don't deal with massive compute problems -- just how night-and-day the difference is between CUDA and its alternatives."
"Because NVIDIA controls CUDA and their GPUs, they control the entire stack. Software, firmware, and hardware. This is the same thing that gives Apple a killer advantage on their platforms. NVIDIA has the same thing, and even aside from programmer productivity/ease (which, as we've learned over the years, is itself huge), stuff in CUDA just magically works much better. Behind the scenes, CUDA performs significant optimizations."
AMD has its own "drop-in" replacement for CUDA called ROCm, but "Even if you were an ambitious PhD in computer science who is also an outstanding programmer who somehow wants to write a new AI-related library in OpenCL or Vulcan or whatever, you still wouldn't get the performance you could with CUDA (at least not without a massive amount of extra low-level programming effort)."
