"It just isn't conceivable that you can design a program strong enough to beat players like me."

December 16, 2015

Arm7 vs Arm5: How many more ELO?

By default, it's expected that an arm7 compile plays better chess than the same source compiled for arm5. We know that arm5 is older, standard, trustable and compatible while arm7 intruduces new commands that increase the performance and arm-8 goes even further to cover the requirements of an 64-bit Android.

Shortly said: arm5, 7, 8 are something like Intel's MMX, SSE, AVX and so on. The progress does not and will not stop.

Regarding the engine compiles for Android, Jim Ablett is recently replacing some of his older arm5 compiles by arm7 but previous arm7 introductions i've done for Rapidroid ranking had not yield remarkable ELO improvements for Cheng and Deep Saros.

Since i don't like practices conflicting with theory, i wanted to insist on my humble assumption of +30 to +50 ELO between two target architectures. So, i've just conducted a self play match between two binaries of Scorpio 2.7.7, using the last 50 openings of TCEC-7.

Both versions played with default settings, without scorpio.ini installed under Chess for Android.

At 5 seconds/move fixed, the Arm7 compile won the match with 57,5 to 42,5. The ELO gap is 52 ELO in these conditions, not far from expected, taking into account the usual self play distortion, shorter time controls and the error margin of only 100 games.

Anyway, theory meets practice. Now, i can sleep happier.

No comments: