"Worse is that it has performance/power tradeoffs, see here.""

Yeah, throttling and slow memory currently negates effects of AVX512. Skylake-SP and X has slow L3 as well, little bit faster then RAM. I guess that with DDR5 situation will change,

Future CPUs will have different capabilities. If you use AVX512 code today, and are disappointed with the lack of performance improvement (if any), then you might find that tomorrow with newer CPUs/mobos/DRAMs it gives fantastic improvements with the same code. So I suggest developing the code now even though it appears to be useless (or nearly so) and prepare to reap the benefits later. You can be the first to have your code run at breakneck speeds when everyone else is still using old code.

@revolution: I don't know if that will happen since Moore's Law is basically dead and we're almost at the limits of transistor size. This is a case of power consumption (and heat generation), which requires smaller transistors. In fact that's the only thing that can increase efficiency. I mean yeah, AVX-512 might end up ok to use (without such "penalties") but I doubt much beyond that (AVX-1024?), tbh.

Remember that parallelization (e.g. stacking transistors in 3D, on top of each other) does not increase efficiency. Doubling the number of transistors at the same size (e.g. 2 layers), doubles the power consumption and performance (assuming 100% parallel), so efficiency is the exact same.

You can decrease the performance-per-core (slower speed) but that's the throttling which we wanted to avoid. Most people use the CPU for single-threaded performance, otherwise we'd just be using GPUs.

Especially now that we have easy GPU computing with Vulkan, no more vendor-specific crap.

FYI, in case anyone is curious, the site/article is back up. Undocumented RDTSC (OS/2 Museum).

Old, pre-existing code might run slower or even break entirely on newer OSes, compilers, cpus, etc. So even what you have isn't guaranteed to remain working!

Wasting time on new cpu extensions (or compiler standards) in the blind hope of "better! faster!" is overly optimistic. Sure, you can take a look, try to add it, but it might actually perform worse!

Regressions, downgrades, deprecation ... bah.