I mean yeah, you can split a lot of hairs there. But even so, even if you assume the RPi 4 can be optimized to be 20% faster than the Atom across the board, it's still the same ballpark. In that ballpark mainstream amd64 support is a big advantage.
Re:Rock Pi X Intel Cherry Trail Board
by vardump on Thursday September 12, 2019 @10:34AM
(#20951906)
NEON optimizations can have pretty large effect (up to ~4x), just like SSE. The benchmarks with the largest differences are suspect, especially SGEMM.
Also don't underestimate how much RPi4 throttles under multicore load. Even the cheapest $1 heatsink will make a big difference.
In general I agree, though. It'll still be in the same ballpark.
Another issue that can skew benchmarks is library optimization situation. On x86, most things are SSE/AVX optimized, but on ARM side NEON optimizations can be slow implementations or even completely missing.
AArch64 ISA has changes that eliminate some false dependency chains (more general purpose registers and NEON register layout changes). This blocks out-of-order engine from making optimal use of CPU resources. When you hit this issue, 64-bit ARM code can run quite a bit faster.
Raspberry Pi's weak spot has always been rather slow memory bandwidth. RPi4 improves this (DDR4 instead of DDR2), but it's still the weakest spot. Just 32-bits wide bus to SDRAM.