3 ms·

As Apple and AMD are currently clearly demonstrating, SIMD just really doesn't matter much. Only a portion of the workloads that are commonly used can be profi

As Apple and AMD are currently clearly demonstrating, SIMD just really doesn't matter much. 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.

monocasa 6y ago
The thing with SVE though is that it's one of those CDC6600 inspired designs like RV-V and ARM Helium that does a much better job abstracting the number of hardware vector lanes. That means way better power consumption at the low end transparently which is very much on Apple's radar.
TinkersW 6y ago
Your example fails to point out why AMD and Apple are able to compete despite having smaller vector widths, and no it isn't because "SIMD just really doesn't matter". 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.