Is geForce 1050 good enough for Oakfoam GPU NG06 version?

[Hong Lu]

Hi Detlef, 

I just found out that it is possible to upgrade my old legacy box  HP dc7700  PC box with a GPU like  EVGA Nvidia GeForce GTX 1050 SC Gaming 2gb, 

which cost only around 100 bucks from Ebay. 

My real concern is,  is GPU version of NG06  Oakfoam setup feasible with such a cheap setup?   Feasible I means that within 1 minute per move limit  

oriented toward DGS like turn based go games, with full potential of NG06  deep learning needs. 

I know my current  CPU-only setup at DGS is really not an ideal thing.  In my own testing,  CPU-only setup requires more than 5 minutes calculation per move in many cases. 

I am amazed by the 1 dan strength at DGS for the oakfoambot with only 1 minute per move setup.  How much of difference should I expect if I upgrade into geForce 1050 setup?

Thanks, 

Hong

[Detlef Schmicker]

From wikipedia I see that the 1050 has 1862 GFlops my gtx 970 has 3920 GFlops.

so you should be about half as fast as my version. My hardware play ng-06 move in about 5s, you should do it in 10s, if everything scales fine. (both cpu and gpu should be about half as fast...) If your cpu is relatively slower you might change expand_after a little without loosing too much strength....

Detlef

--
You received this message because you are subscribed to the Google Groups "Oakfoam" group.
To unsubscribe from this group and stop receiving emails from it, send an email to oakfoam+u...@googlegroups.com.
To post to this group, send email to oak...@googlegroups.com.
Visit this group at https://groups.google.com/group/oakfoam.
For more options, visit https://groups.google.com/d/optout.

[Hong Lu]

Sounds very good,   my upper limit is 1 minute per move for turn based go games.   That means with geForce 1050 should have full strength of NG06 at DGS. 

Based on your comments that NG06 should be KGS 5d,  at DGS   it should be around 3d to 4d.    

I am pulling the trigger to move forward with my upgrade plan.

Thanks, 

Hong

[Detlef Schmicker]

the 5d statement comes from cgos strength and a statement from aya author ...

but it was for ng-04 

ng-06 is around 200Elo stronger on CGOS, should even be 6d on kgs (close to hirabot, as hirabot is a only little stronger on cgos)?!

I still think you have probably a random bug, should be stronger than 1d kgs with your setup?! But maybe solved by gpu anyway, so we will see :)


I hope to play march slow bot tournament on kgs (I am training a 20 layer net at the moment and am not sure, if it is ready by the tournament), and if everything works fine play two more days to get a rank on kgs .... At the moment I do not have resources to play, need all for training.

Detlef

[Hong Lu]

Ok ,  KGS 6d is equivalent of DGS  4d to 5d.

The administrator over at DGS thinks that there are 2 stones difference between DGS ranking and KGS ranking.  I tend to believe the ranking difference is only at 1 stone.

Hong

[Hong Lu]

Just let everybody know that my upgrade of Oakfoam NG06  with 1 GPU has been successful. 

It took me some effort to make it work because it appears that in order to compile  Oakfoam GPU version of NG06 successfully,  the caffe library has to be linked at  /usr/local/lib .   

The compiling did not work if caffe is enabled  under   /etc/ld.so.conf.d without symbolic link at /usr/local/lib .   The configure  script would simply complain that "caffe" not found even though I supplied the correct path. 

The new GPU version of NG06 is running at DGS at:   http://www.dragongoserver.net/userinfo.php?uid=97704

[Detlef Schmicker]

great to hear you got it work. configure scripts are awful, especially in case of components not standard on distributions as caffe :(

Thanks for running it on dragongoserver. I had a look at the info. It says it supports japanese rules. I dont know what your script does, but at the moment oakfoam only supports Chinese rules.

To pseudo support Japanese rules one must at least count the passes and correct the result with them.... 

I have some days off, perhaps I will work on passing out the game. Since DCNN support this does not work any more (as unprunning did the job before and now only considering cnn results)

Detlef

[Hong Lu]

Yes,  the scoring is handled by DGSlib using gnugo method,    Oakfoam only need to do pass and rest of the Japaese rule are handled by other software. 

In fact,  the DGS  oakfoam bot is handled such as if other party issue "PASS",  Oakfoam Bot will follow with PASS as welling going into final scoring

Hong

[Hong Lu]

new Oakfoam Bot (NG06) with 1 GPU at DGS rose to 4 dan now:

http://www.dragongoserver.net/userinfo.php?uid=97704

[Detlef Schmicker]

Thanks again for running oakfoam on DGS, good to see it is playing ok in real world:)

By the way, was the estimation of about 10s per move on your machine correct?

Detlef

Am 02.03.2017 um 03:08 schrieb Hong Lu <hong...@gmail.com>:

new Oakfoam Bot (NG06) with 1 GPU at DGS rose to 4 dan now:

http://www.dragongoserver.net/userinfo.php?uid=97704

--

[Hong Lu]

Double that amount in my machine. 

I manually checked the per move timing and think it takes roughly 20 seconds per move most of time.   But we have to consider the extra time several seconds needed to load data and memory because the set up is to quit 

GTP after each move and restart from beginning move to genmove for each new move,  the real  per move time may be around 17 seconds 

Hong