Re:Apple Hires Key Chip Designer From ARM as Own Efforts Ram...
by tatersolid on Wednesday June 26, 2019 @11:44PM
(#20291595)
There are so many different flavors of AMD64 that it’s not even a single target from my perspective. Check out the quantity of compiler flags and ifdef statements you see in just about everything.
I’ve seen binaries in production run at 30% of the speed seen DEV/QA. These would also segfault every few hours. But debug builds ran fine in production.
Root cause in that case was the Intel server had some hardware instructions available that the Intel developer and QA systems didn’t. The runtime chose to use these fancy new instructions, which turned out to be a buggy and slow code path.
Again, if you’re not doing DEV/QA on hardware (and OS, dependency, and firmware versions) that differ from production you will eventually be bitten.
Everything is so complex and there are so many layers involved it is impossible to have confidence. Containers make this worse in my opinion by adding another complex (and buggy) abstraction layer.
Sure, but when it comes to platform enablement, that only really needs to be done once per vertical/native dependency. AArch64 has all of those things you listed. Don't know about .NET (though there's not that much .NET development on Linux, at least not when compared with the amount of Node and Java), but OpenJDK, Node, C/C++/Rust/Go toolchains all have mature support for AArch64, among others.
It's obviously less mature than the 40-year-old x86 ecosystem, but it's basically all there. Furthermore, the ease with which new ISAs can be adopted by toolchains and libraries is accelerating, so POWER and RISC-V are also getting there in terms of software availability.