Following the decision to reinvestigate whether SF5 plays better than 121014 compile, i have interrupted round 20 of the Rapidroid experiment and started a new version of the duel on my Galaxy Note II, using once again all 32 openings of TCEC-6 superfinal.
In contrast to what happened on Intel Z3745D and in parallel to bullet results reported by Kai Laskos at talkchess forums, my first results at 15 sec/move confirm 121014 overcomes SF5 when played on Exynos cpu, using arm7 compiles.
After 64 games of which 42 were drawn, Stockfish 121014 leads 13 - 9 vs Stockfish 5, counting only the wins.
This sums up a 34-30 score, placing 121014 22 ELO higher than SF5, using the famous formula like 400 x LOG (34/30)
I've also checked again the average nodes of both versions on two different cpu's:
Stockfish 121014 / arm7 / Exynos 4412: 802 kNps
Stockfish 121014 / i686 / Z3745D: 770 kNps
Stockfish 5 / arm7 / Exynos 4412: 718 kNps
Stockfish 5 / i686 / Z3745D: 770 kNps
What's interesting is that on Exynos, the 121014 had 12% more node speed than SF5. This gap totally disappears on Intel and they both hit the same score, just like they were hitting a maximum at 770K. Weird but true...
Different behaviour on Intel compared to Exynos explains half of the difference. I believe increasing the games played on both cpu's might help fully reveal the truth.
For the moment, i can only extend the verdict like:
Stockfish 121014 is the strongest engine available on Android for arm7 cpu's.
Stockfish 5 is the strongest engine available on Android for intel cpu's.
All findings to update after playing more games.