It's a bit odd that this would happen so soon after Cloudflare started testing & praising ARM-based server (and specifically Qualcomm's Centriq). For reference:
https://blog.cloudflare.com/arm-takes-wing/ where they find Centriq is mostly performance-competitive with Skylake at a significantly lower TDP/power point modulo platform support/optimisation e.g. Go's ARMv8 backend being immature and lacking optimised assembly routines
https://blog.cloudflare.com/neon-is-the-new-black/ where they demonstrate the latter by simd-optimizing jpegtrans for ARMv8[0] leading the Centriq 2452 to reach (and even overtake) Xeon 4116's performance-per-worker and blowing its throughput-per-watt out the water (25 image/second/watt on Centriq 2452 versus under 10 on the 4116)
[0] as they'd previously done for AMD64, though starting with equivalent optimisations using SIMD intrinsics then adding ARM-specific assembly for further gains as "the compiler in that case produces somewhat suboptimal code" using intrinsics[1]
[1] because according to a commenter and SO[2] GCC inserts unnecessary register copies when involving NEON yielding to very poor machine code, Microsoft's ARM compiler is apparently stellar and Clang used to be meh but greatly caught up between 2012 and now
https://blog.cloudflare.com/arm-takes-wing/ where they find Centriq is mostly performance-competitive with Skylake at a significantly lower TDP/power point modulo platform support/optimisation e.g. Go's ARMv8 backend being immature and lacking optimised assembly routines
https://blog.cloudflare.com/neon-is-the-new-black/ where they demonstrate the latter by simd-optimizing jpegtrans for ARMv8[0] leading the Centriq 2452 to reach (and even overtake) Xeon 4116's performance-per-worker and blowing its throughput-per-watt out the water (25 image/second/watt on Centriq 2452 versus under 10 on the 4116)
[0] as they'd previously done for AMD64, though starting with equivalent optimisations using SIMD intrinsics then adding ARM-specific assembly for further gains as "the compiler in that case produces somewhat suboptimal code" using intrinsics[1]
[1] because according to a commenter and SO[2] GCC inserts unnecessary register copies when involving NEON yielding to very poor machine code, Microsoft's ARM compiler is apparently stellar and Clang used to be meh but greatly caught up between 2012 and now
[2] https://stackoverflow.com/questions/9828567/arm-neon-intrins...