2018-11-23, 18:10
(2018-11-23, 04:24)J_E_F_F Wrote:(2018-11-22, 12:32)Milhouse Wrote: the ARMv8 performance benefit is not significant (from past testing, 10-15% or less compared with ARMv7),significant is such a subjective term. I'd call a 10-15% gain from software with no change in hardware very significant.
People pay hundreds on a new phone, or thousands on a new computer for a 10-15% hardware CPU gain. To get that from nothing more than software on existing hardware, is truly significant.
We'd love to achieve a 10-15% performance gain from a simple compilation switch, if only it were that simple!
Adding a new ARMv8 RPi architecture would mean supporting a completely new device ("RPi3" in addition to the existing RPi ARMv6 and RPi2 ARMv7 devices) with a dedicated RPi3 ARMv8 build and associated RPi3 ARMv8 binary addons. That is an enormous challenge for a team that operates solely on donations and volunteer effort. And that's without considering all the issues that come with a 100% ARMv8 64-bit (aarch64) build, which basically means it's not currently possible without breaking a lot of required stuff, so the best that could be achieved is a hybrid 64-bit kernel/32-bit userland, which isn't really that great as a choice, or a significant move forward. It would be creating the ARMv8 build for the sake of it, generating a lot more work, without achieving anything significant or worthwhile.
When testing ARMv8 over 2-3 years ago I only tested ffmpeg performance, which was probably 32-bit. I don't actually remember the precise numbers now, which is why is said "10-15% or less" but I would emphasise the "or less" part as the gains were small.
So yes, it is entirely subjective but given the choice of supporting what would effectively be a new device on an immature arch with a very modest performance gain - all of which requires more compute time for build cycles, more storage, and more forum discussion about ARMv8 specific bugs/issues - or recommending a small ARM overclock, the choice was obvious!