Why the fuck is none using Intel MKL for AI or video codec despite being all matrix multiplication in disguise?

At least they admit that parts of it are optimized for Intel processors.

Back in late 90s or early 2000s there was a scandal where an Intel math library was optimized for only Intel chips. But benchmarks were using the library and saying AMD sucked because of these math routines.

Not sure what math libraries they use for benchmarks these days.

What do you mean? MKL is being used for AI, just not by default because CUDA is king still and MKL barely makes a difference on most consumer chips

You're referring to the old issue where Intel's compiler and math libraries favored Intel CPUs by detecting the vendor string. If it wasn't "GenuineIntel", even compatible AMD CPUs were given slower code paths. This made benchmarks unfairly favor Intel in the late '90s and early 2000s.

Today, many libraries are more vendor-neutral. OpenBLAS, BLIS, and Eigen are commonly used and support both Intel and AMD. Intel’s MKL has improved, but still may favor Intel chips unless patched.

For fair benchmarking now, it's best to use open-source math libraries and vendor-neutral compilers like GCC or Clang, and avoid relying solely on proprietary tools like Intel’s compiler without checking how they handle CPU dispatching.