It's even harder than that. SVE vs NEON performance will also hugely depend on the vector length the given algorithm requires, and the stress that the instruct

It's even harder than that. SVE vs NEON performance will also hugely depend on the vector length the given algorithm requires, and the stress that the instructions put on the memory subsystem. Memory hierarchy varies by product and will likely continue to do so regardless of what M2 does. In the end, I echo eyesee's comment, for best performance one should really use Accelerate.framework on Apple's hardware.