Why not post the home page here? It is open source right? I'm not so familiar with the dragon server and was lost in finding any info.

Here is source link:
https://github.com/zakki/Ray

Thanks. I'll take a look.

Thanks for the link, I see one 6 dan DGS user is 7d on KGS.

@Shoreline
It seems that Rayon runs on cgos with 4 cpu's but no gpu (rn.3.9-4c until 2/19 and rn.3.10-4c since then). Do you have any idea how much impact the gpu makes?

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Rn.3.3.1-4c0g doesn't use gpu. http://www.yss-aya.com/cgos/19x19/cross ... -4c0g.html
Other Rn.xx-4c uses i7-2600K (or 6700K) + GTX980ti

I think GPU impact on RN3.9 is quite large. Rayon seems to be able to take advantage of GPU resource fully in my setting.

In DGS turn based go server set up, GPU may not matter because you can make it longer thinking time, say instead of 10 secondes per move, you can strengthen the bot by making 20 seconds or even 1 minute per move. However, in KGS or CGOS, I suppose the rule is much tighter, because the bot may lose by time out.

I took back my words. the difference with GPU is not just "large", it is dramatic. I was using same playout number in CPU only mode
for a couple games, the RN3.9 (CPU only) is quite weak, no where near the 1 GPU version I set up at DGS.
It could be that CGOS was set up by author with 4 CPU who knows how to optimize the parameters. For me with no real understanding of the software, the difference is dramatic.

Were you able to get this compiled for windows? I used MWIN2 with gcc 6.3.0 and CNTK beta 11 but I get tons of segmentation faults. It does rarely run, although even when it does run it just exists to desktop. No interactive gtp prompt like I expected.

I tried using the CNTK "cpu only" version since I don't have an NVidia GPU. Just to see if it'd work.

When I compile this it says it is version 8.0.1, the same as a compile from one year ago.

There is link "6 releases", pick the latest one, rn3.10, which was released just couple of days ago

Oh there's binaries too, at least for 3.3. I was trying to compile myself direct off the github.

I tried compiling

https://github.com/zakki/Ray/releases/tag/rn-3.10.2

for Windows with MWIN2 and gcc 6.3.0. To get past compiler errors in UctSearch.cpp I had to change "std::atomic_int64_t" to "std::atomic<std::int64_t>" in the appropriate spots, and also change atomic_fetch_add in some places to be "(long int)1) instead of "1)" at the end of the function all.

Unfortunately, it will seg faults a lot if I try running it after compilation. This is with the CNTK beta 11 CPU-only version. The last binary release was rn-3.3.

I assume you are one of the developers. Is there any chance you could add binaries for 3.10?

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Just report to everyone here that Rayon Bot (RN3.9) at DGS achieved 6 dan level and slaughtered a human 5 dan player on an even game. I actually
do not know exactly the upper limit strength of this Rayon Bot at DGS, could be a lot stronger than 6 dan.

http://www.yss-aya.com/cgos/19x19/bayes.html

Leela 0.9.1 is out in the results. It appears that new leela is significantly weaker than RN3.9, The difference could be several stones.

Just got information from Rayon authoer that RN3.9-4c in the above link CGOS ranking uses 4core i7 and GTX-980Ti GPU. This setting is significantly beefer setting than my DGS Rayon Bot setup

Hard to compare between different program because of different hardware (4 cores vs. 1 core + 1 gpu)

I think you misread, it is not Leela 0.9.1, it is Leela 0.9 running on a single core. It's stronger than Rn3.8 on 4 cores + GPU.

Leela0.9.1c1g in the link. I suppose that is leela 0.9 1 CPU with 1 GPU. In the go AI, GPU calculation is the main calculator,
CPU does not matter that much. That is still a comparable hardware.

Just a note --- if we are going to discuss the effect of hardware on the playing strength of these programs we can't just say things like "a 4 core I7". That encompasses a range in crunch power almost 4:1 between the strongest and the weakest I7 << about the same ratio as between the strongest CPU's being used by go programs and the strongest I7's >>

Leela makes great use of the cpu though. GPU is only used for the neural network related computations, the rest of the program is all cpu based.

Yes GPU is important and no GPU means weaker play, but a strong cpu will make a much stronger Leela.