20:34 <hexdump0815> maybe someone has an idea regarding the following: if i run an application which mostly doing dsp style and fpu heavy code (vcvrack - a modular synthesizer running purely in software) on both an s905x2 (1.8ghz) and an s905x3 (2ghz) it is using about 20% (relative) less cpu on the s905x2 than on the s905x3

20:42 <hexdump0815> so looks like it runs less cpu efficient on the x3 than on the x2 (actually it is not running any faster on the s905x3 at 2ghz even than on an old s905 or s905x at 1.5ghz) which is surprising me quite a bit (especially keeping in mind that the x3 is even clocked higher than the x2)

20:43 <hexdump0815> i did a compilation test on both the x2 and the x3 and there the difference was only 5% between them (compiling openssl on all cores) - those 5% might be due to differences in memory, sd-card etc.

20:46 <hexdump0815> are there any known errata for a55 or is it maybe that the a55 cores are worse at some things than a53 cores (like fpu stuff or certain cache intensive operations)? i verified it across several different systems (even an odroid c4 with the amlogic 4.9 kernel) and for this kind of workload the s905x3 is simply not any faster than the s905 in an

20:46 <hexdump0815> odroid c2

20:49 <hexdump0815> or is there maybe something not yet perfect on the linux kernel side when initializing the a55 cores in the s905x3 - i guess not (and would even not understand the details) but wanted to ask it here ... an allwinner h6 with a53 cores at 1.8ghz is about equal percormance wise to the s905x2 ... so looks like really the a55 based s905x3 is somehow

20:50 <hexdump0815> standing out here

21:12 <hexdump0815> one thing which might be special about that code: it uses 'asm volatile("yield" ::: "memory");' as a replacement for '__mm_pause();' in intel code - can this maybe make any difference for a55 vs. a53?

21:14 <hexdump0815> i also tried to compile with gcc8 tuned for a55 vs. a53 before, but this also did not make any difference (all this is on ubuntu 18.04)