Zakki would you be willing to release a binary version linked to the CNTK CPU Only library as well for those of us with AMD cards so we can try Ray out? (Those of us who can't get it to compile. )

Thanks!

Edit: Whoops! I was reading it that I had to have an nVidia GPU to even run the program. It works fine without a GPU, just much much slower. Thanks again.

Can Ray run uses more CPU than GPU?
My computer is i7 -4790 with NVIDIA GeForce GTX 745.
Run test for it with Leela, and the score is 1-1 right now. Leela 0.9 completely destroyed it the second game.

Ray 3.10 has won 24 out of 31 games vs Leela 0.9.1.

http://www.yss-aya.com/cgos/19x19/cross/Rn.3.10-4c.html

I imagine you'll see Ray pull ahead as you run more tests. What program are you using to play them again each other? GoMill? Or by hand?

I did it by hand with default setting. Checking the GTP see that it only run up to 10000 po per move in 10 second.

Could you please explain how to run it without GPU? So far I have downloaded three missing dll's for VC, but still the program stops working at the start.

All the DLLs I needed were in the release zip file. Just run the exe from the directory all the DLLs are in.

This is the 3rd game.

Both AI used 10s/ moves. The only command I used for Ray was --pondering. The rest is default.

This game Leela was leading early, then she made mistake in the upper side. Lead to Ray slowly lead.

White = Rayon
Black = Leela
Komi 6.5
W+Res

Rayon Bot lost a few games in DGS at 6 dan. I think it is now reasonable to say the strength of this setup (2 cpu, 1 GPU) on 50000
play out mode is roughly at DGS 6 dan. On seconds per move calculation, I checked a few moves, more or less, it will move in around 20 seconds.

Hey, I see Rayon can organize its timing with "--time" flag. Can you implement this as GTP commands, so we can use the GUIs (Drago here) interface for setting time controls? Thx for the engine.

A comparison Leela/Rayon, how can matches be fair? I'd like to have them calculate a fixed time per move, but Leela has no flag for this. Both have an argument for # of playouts, but Leela needs much more time for calculation if both do same # of playouts...

How about "time_left" GTP command as workaround?
Ray ignores "time_settings", but it understands "time_left".

Leela has full support for time_settings and kgs-time_settings. I'd suggest setting up something simple like:

time_setting 0 30 1

That tells the go engine you are using Canadian time, have 0 main time, 30 seconds per period, and need to play 1 stone per period. This would make Leela think at most 30 seconds per move.

You can set the playouts for both engines different to make them use roughly the same amount of time. I would play some games and look how much time they used in average for the whoole game. As time and playouts is likely a linear function you can calculate the number of playouts for an average time per move or per game for both engines.

I think I don't have enough GPU to make this bot as strong as the one on server. For some reason, Ray lost most of its game to Leela on my computer.

I only use GeForce 750 so maybe that's the limited. Even with the i7 4970 won't help.

You could simulate a stronger gpu by letting Ray have more thinking time than you give Leela.

Not sure this will work. In my CPU only testing, on the same number of play outs (50,000 play out), the ray on CPU only is a lot weaker than GPU version.

Ray RN3.9 or RN3.10 strength appears to be highly depending on the GPU quality.

Dependency on GPU quality is not new. Oakfoam Bot (NG06) also has this issue. NG06 CPU only version in my testing is weaker than GPU version.
But the degradation at NG06 is a lot better than RN3.9, not as dramatic as ray

Interesting. I assumed cpu version would execute same code, just slower.

Does Rayon support the 45% of all users which have an amd gpu?

Can I use 4x xeon cpus with Rayon?
And 2-4 gpus?

Does it work in the Smart Go or Leela gui for windows?


Can someone post here the exe?
No matter if this would be 10% slower on my machine.

Binaries: https://github.com/zakki/Ray/releases

Yes it works in Smartgo. Use the non GPU binary. EDIT: There is no such binary. I guess as an AMD user you cannot use it.

The GPU version will work, it just runs in cpu only mode. And it's 30x-100x times slow per author without GPU.