[Computer-go] New AMD processors

["Ingo Althöfer"]

Hi,

AMD has published a new (fast and cool) processor, the Ryzen.
Did some go programmers already collect experiences with it?
Do they combine well with GPUs?

Ingo.
_______________________________________________
Computer-go mailing list
Compu...@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Gian-Carlo Pascutto]

On 03-03-17 21:29, "Ingo Althöfer" wrote:
> Hi,
>
> AMD has published a new (fast and cool) processor, the Ryzen.
> Did some go programmers already collect experiences with it?
> Do they combine well with GPUs?

I'm not getting one until there are mainboard reviews out, because there
seem to be early BIOS issues, but some things can be predicted:

- *without* GPU: Ryzen has half the AVX power, so running the DCNN on
the CPU goes at half speed. But this is not the only thing the program
needs to do, and you do get twice the amount of cores. So I think it
should easily make up for the difference, compared to the 4 core Intel
chips.

- *with* GPU: in this case only integer performance matters, which
appears to be outstanding. e.g.:
http://www.hardware.fr/articles/956-14/ia-echecs-stockfish-komodo.html

Price/performance should be outstanding.

--
GCP

[Detlef Schmicker]

I looked into this too:

oakfoam would not benefit a lot from more cpu power at the moment, with
4 cores I mix 10 playouts with the value net in the ratio (3:7) at the
moment.

In case of buying a Ryzen: take care the board allows two GTX1080 Ti
(wait till end of march to buy them) and buy a power supply with 1kw
(and send a copy of the machine to me) :)

Detlef

P.S.: oakfoam status
http://www.dragongoserver.net/userinfo.php?uid=97704
http://www.yss-aya.com/cgos/19x19/cross/NG-06.html

[Álvaro Begué]

Oh, you are using a value net? How did you train it? I don't see anything about it in the bitbucket repository...

Álvaro.

P.S.- Sorry about the thread hijacking, everyone.

[Detlef Schmicker]

yes, it is trained from 4d+ kgs games (about a million I think). I am
not sure if I put the actual training scripts into my branch right now.
And I did not document how to combine both nets after training, but it
is straight forward, caffe allows to load multiple caffemodel files and
saves a new one :)

at the moment it is in my branch:
https://bitbucket.org/dsmic/oakfoam

I use a combined policy value net, a call gives the move prediction and
the value. The full net and configuration for NG-06 (10k playouts, about
5s per move on i7-4790k gtx 970) is in the branch (you can not run it
nicely on kgs at the moment, it is missing passing if opponent passes)

Detlef