I dont much like this article since SIMD is not “niche” as the author says, even if people aren’t aware of it. It’s hard to pick up a single piece of software and say that “yes my overall workload overall will improve” by turning it on, but if you are aware of it, it will speed up most non-branching loop operations via auto-vectorisation.
Having had to work on software that did runtime dispatch for SIMD - there is a small performance cost to this. For most software it’s just not worth it. Compiling multiple versions is preferable but more confusing for users, and more costly for the developer in CI and build time. There are therefore good reasons not to support multiple microarchitectures and picking one released 12 years ago is a good compromise. I wouldn’t at all advocate picking AVX512 but this is not that - almost all consumer and professional grade hardware sold in the last 10 years supports AVX, and it’s unreasonable to expect vendors to continue supporting an ever shrinking niche…
“Runtime dispatch” does not require that the CPU feature be queried on each call. There are designs that can do a single query on launch and set up the environment so that future calls do not incur any additional penalty.
Having had to work on software that did runtime dispatch for SIMD - there is a small performance cost to this. For most software it’s just not worth it. Compiling multiple versions is preferable but more confusing for users, and more costly for the developer in CI and build time. There are therefore good reasons not to support multiple microarchitectures and picking one released 12 years ago is a good compromise. I wouldn’t at all advocate picking AVX512 but this is not that - almost all consumer and professional grade hardware sold in the last 10 years supports AVX, and it’s unreasonable to expect vendors to continue supporting an ever shrinking niche…