January 16, 2016

Android UCI engine update: Ivanhoe 9.47c Beta

The second nice surprise which came in the engine package sent by Valery Kotelnikov is an alternative build of Ivanhoe that i've been searching all over the web without success since long ago.

It wouldn't be a crucial requirement if the build 9.46h could perform as expected, precisely said, if it could use hash memory. Although the engine runs okay, it has issues related to the way it was compiled:
* Uses no hash memory at all regardless of the setting
* Uses all available cores or crashes if the user sets another value

Given that Rapidroid ELO of Ivanhoe is only ~2700 vs 2900+ expected, any other alternative is welcome. Now that 9.47c Beta is in the pocket, i can check whether this build works better than 9.46h.

And interested users can choose to give a try before me: HERE

28 comments:

  1. Я когда-то этот движок тестировал,хорошо играет,думаю в Rapidroid в топ-10 войдет.Движок очень "цепкий",любит эндшпиль.

    ReplyDelete
  2. I tried it, hash seems to be used, but cannot choose cores and only 1 core is running...And Toga using only 2 cores though I chose 4.

    ReplyDelete
    Replies
    1. TogaReturns uses 4/4 cores with RK3188. It has sern 1023 knps in 60 seconds test. I guess 1200 with exynos 4412. I can't do it now because both devices are running tourneys.

      Delete
  3. I forgot to add - Samsung Galaxy s3, exynos 4412, arm7, LP 5.1.1.

    ReplyDelete
  4. Ivanhoe 9.47c одноядерный.

    ReplyDelete
  5. Ok, i thought it's supposed to run on 4 cores, like IvanHoe 9.46b.

    ReplyDelete
  6. Thanks for testing. On RK3188 quad core the result is the same for me. Unlucky Ivanhoe can't get together hash and cores. Now i'll try to conclude with a gauntlet to see which feature has more effect on the strength.

    ReplyDelete
  7. Gurcan, by the way - what temperature limits for CPU and battery you think are safe for long runs?
    I downclocked CPU to 1.2GHz (from 1.4GHz), and I got at max 65 degrees for Cpu and 48 degrees for battery.From what I know it should be acceptable?
    Also waiting for your Toga test on exynos 4412.

    ReplyDelete
    Replies
    1. Below 70C is safe for Exynos 4412. Monitor the frequency in parallel. The risk is throttling. Exynos does not freeze! Games will then continue at 1.0 ghz and finish. Don't forget to check my previous posts containing "throttling" and use the monitoring tools i've recommended.

      Delete
  8. With two multi thread engines playing and without special prevention exynos throttles all the time. Beware...

    ReplyDelete
  9. I'm using monitoring tools since long time ago, beside CPU Stats I'm also using Cool Tool, which is also a godsent tool for geeks ;)
    I do not have problems with throttling as I'm using custom OS, but I remember this pain of CPU freq going down if even battery went over 42 degrees :)

    I've also started a mini tour 300+2 (first Silversuite openings), between "old" Ivanhoe and "new" Ivanhoe, as I know you're little busy. I'll wait until at least 20 games will finish and I'll inform about the result.

    ReplyDelete
    Replies
    1. Threads matter much more than hash according to what i'm observing right now: 9.5-26.5 after 36 games out of 612. 180+1 using Tcec-7 opening suite on RK3188 @1.2 ghz. Depths reached show the gap. I'll continue to follow that gauntlet but it's already obvious.

      Delete
  10. That's weird, because I'm having quite opposite results with *new* 1 threaded Ivanhoe 9.47 with 128 hash (sorry Alexei ;)), the score is +9-0=4 for the *new* one (?).
    As I wrote in post above - 300+2, Silversuite openings.
    It looks like shorter time controls favours the 9.46 then, and more time = more hash meaning for 9.47.
    Interesting :)

    ReplyDelete
  11. Samsung Galaxy s3 1.4GHz 4 cores downclocked to 1.2GHz, custom CyanogenMod 12.1 (Lollipop 5.1.1).

    ReplyDelete
    Replies
    1. Can you please make a benchmark using the following settings for both and report the nps:
      180 sec/move in CfA, starting position, hash 256, 4 cores with 9.46. Set tablebase mem to none and tablebases to none. Thanks.

      Delete
  12. v9.46 - depth 21, 406kNps
    v9.47 - depth 19, 122kNps

    ReplyDelete
    Replies
    1. Ok. According these values 9.46 should easily crush 9.47.

      Delete
  13. I'm curious about feedback from Lollipop users. I fear Lollipop because it restricts the 32 bit engines compiled for arm7 without nopie option. Is there anyone who run such old builds of engines like Critter, Gull 1.2 etc under Lollipop with CfA?

    ReplyDelete
  14. I assume you're running KitKat?
    I ran couple of games with your TC - 180+1, and 9.47 still won't give a chance to win for 9.46, so assumption is that engines without nopie option will perform better on 1 core than on 4.
    I'll run both engines on 1 core for comparison.

    ReplyDelete
  15. Edit: Ivanhoe 9.46 won't run on 1 core on my device.

    ReplyDelete
    Replies
    1. Can you post the md5 checksum of 9.46 you are running? I can run it single core at same nps as 9.47.

      Delete
  16. К сожалению не могу провести серьезного тестирования.Однако если на длинном контроле 9.47c и выглядит хуже,то при 1 секунде на ход результат 30 партий(TCEC-7:1-30,256mb hash,4-core,1.5GHz,Tegra3,OS4.0.4):9.47c-18 1/2--9.46b-11 1/2 .Вот партии:https://cloud.mail.ru/public/AVVy/8SM57uQJx Мне казалось ,что и на более длинном контроле результаты будут не на много хуже.

    ReplyDelete
    Replies
    1. I've checked the pgn. Time control is 1/1 and depths under 10...
      Even with multi threads running, it may be hard to kick in for 9.46. Usually SMP speed up is slower than single thread. We don't speak about single to single duels. 9.47 can easily win these with hash.

      Delete
  17. Что-то здесь не так,надо разобраться.Я бы сам этот движок протестировал,но я Cull 3 против komodo 9.2 гоняю.

    ReplyDelete
  18. Если,как я понял ,у 9.46 не работает hash,то получается, что чем меньше глубина перебора тем весомие наличие hash. Думаю, это вполне логично.

    ReplyDelete
  19. Latest standing: 9.46 4 threads no hash leads by 148 to 46 vs 9.47 1 thread 256 mb (tcec-7 openings, 194 of 612 games @ 180+1)

    ReplyDelete
  20. If you are asking me, MD5 of my 9.46 is 12ED94B63342BEC0062C74220DFD121B.

    ReplyDelete