But hey, maybe they just say 'screw it' and remove it anyway.
But hey, maybe they just say 'screw it' and remove it anyway.
Only a portion of the workloads that are commonly used can be profitably vectorized using SIMD. The curiously perverse nature of SIMD is that the wider the vectors, the smaller the proportion of time used by these portions is, and therefore the less you gain from further vector width increases. AMD is currently spanking Intel in almost all practical vector workloads, despite having half the vector width. Apple isn't far off either, despite a quarter of the vector width.
It wouldn't really ever significantly hurt Apple if they just literally never implemented any flavor of SVE. Spending all that engineering effort on improving scalar throughput probably has much better real-world payoff.
It is because AMD and Apple have wider architectures with more vector ports, they can execute 3 or 4 of these instructions per cycle while Intel can only execute 2(or even 1 in some cases with avx512).
AMD has already said they will be adding AVX512 to the next zen, so they apparently think SIMD matters.
Apple will almost certainly implement SVE, they would be stupid to not do so, and they aren't stupid.
I don't think keeping NEON capability in hardware is going to hold back their chips much, so they probably won't be under any pressure to break compatibility with NEON-using apps anytime soon.
And the iPhone 12 won’t be out of support in 3 years.
It used to be the case if you used NumPy on a Mac it would use Accelerate but it no longer does because the BLAS/LAPACK API is so outdated.