This is because very little uses AVX instructions, hasn't it always been better to optimise software for the GPU rather than AVX instructions, even though they are not perfectly interchangeable.
You might be thinking of the most recent AVX-512 extensions which had significant power management complications delaying adoption. Many common programs have optimized SIMD paths because GPU computing is significantly more complicated and the setup overhead makes it slower for uses which don’t process tons of data - one of the interesting points about the M1 is that the unified memory design avoids the overhead of copying data between the CPU and GPU.
No problem - Intel’s marketing hasn’t helped at all. I notice this mostly for cryptography- we do a lot of data validation so spending up SHA is a nice win.
