Hajin Lee (Haylee) 4p vs. AQ this Saturday at 23:00 UTC

For that matter, both 12am and 12pm denote midnight, since both 12 hours to noon and 12 hours past noon… is midnight.

Noon itself is just m (meridis?) or 0m (historical joke right there, would be downright offensive to Romans).

1 Like

@vargos I concur, at low hardware, Leela > AQ by far!

So I’m predicting today yet another Haylee win… :frowning:

I do see the value in testing low end (GTX 970 is not “HIGH END”) hardware against pros, something that hasn’t been done before. But Haylee isn’t the right candidate for that. Each time she has gotten too lucky due to always accepting games in which she knows she is going to win. That at all sportmanlike for example Lee Sedol or Kie Je.


AQ won on a single 970


Thanks everyone for tuning in! ( we reached 200+ viewers). I hope you enjoyed the game! Haylee will be posting a video of her playing this game on her Youtube channel some time.

an idea, zenleebot, next up, AQ vs Dywrin on a gtx 1050 mini (only $120 before the bitcoin craze )

I only have access to a GTX 970/ i5 and a GTX 1080/Xeon E5-1650. If Dywrin want’s to play I’m happy to play on the 970 system but with lower thinking time for ZenLeeBot


Or you can keep using the same setup and invite "hanayeol"
he is very popular Go instructor, about 7D, on youtube twitch and patreon
Also of Korea player like Haylee.

And how about 3P Pro Janice Kim?


I am happy to run the same setup against pros if they wish to play against ZenLeeBot.

With tonights win, I’m going to rent some Nvidia V100’s on Amazon’s AWS and take AQ for a spin…
Their p3.16xlarge instances offer up to 8 x Nvidia V100 each with 5,120 CUDA cores and 14 TFLOPs for total of 40,960 CUDA cores and 112 TeraFLOPS. I read that AQ is can take advantage of graphics processor scaling…

Would be interested to see a rematch with Haylee on this kind of specs.

@ZenLeeBot on a more serious note, if I rent 4x Nvidia V100 on the amazon cloud and set up with AQ would you be interested in doing/hosting either a rematch with Haylee or a new match with some other Pro?

I am not experienced with running Amazon AWS server instances. If it is set up for me or there is a tutorial for me to set it AQ up I can do it. I will need to have a few test runs as well to make sure it works reliably and without fail.

Okay. I usually do the Windows instances since I don’t know Linux. So it would be a Windows Server 2016 instance (basically just like Windows 7/10) and rdp (remote desktop) to it using the elastic amazon aws public IP assigned/associated with the local windows credentials to that instance and the GPU’s will already be attached and all I would have to do is install the latest Nvidia drivers for it, install Sabaki and AQ and that’s pretty much it.

The Nvidia V100 chips costs around $10,000 each, so it is outside of my price range for sure. But Amazon offers to rent them out on-demand, the price ranging from a little over $3 to $24 per hour, (from 1 to 8 x units) and no charge when the instance is stopped. Then all I have to do is boot them up for a game or etc. Each V100 is comparable to a Titan V (but stronger), with a Titan V significantly more powerful than a GTX1080Ti… I would be interested to see just how much more powerful AQ can play with 8 or at least even 4 V100’s.

I’ll PM you once I’ve set it up.

At some point I’ve had the ideas that AWS Lambdas would be perfectly suited for utility AI deployments. They running limit is 5 mins, are self contained and you pay only for the time they run. Unfortunately with Lambda’s you can’t influence your hardware a lot, so that rules out all the GPU hungry engines. I do hope that with AlphaZero architecture Lambdas should be a feasible option.

That “test” wouldn’t help you much if you’re looking to gauge performance/strength gain by hardware. You would have to pair the weaker hardware AQ vs the ostensibly stronger AQ. ZLB confirmed that “better” hardware doesn’t always lead to victory, so you would have to measure performance by having the two play at least 30 games (to be able to make something of the statistics).

The idea of such a test is to run it on 8x V100 and give Haylee 3 handicapped stones headstart. The next version of AQ will deal much better with higher handicaps.


I did a test game, single V100. Leela pretty much got slaughtered… (I found out in order to use 8x V100 I have to get amazon permission so I’ve made a request waiting to hear back. Luckly AQ supports up to max of 8 GPUs)

Update 2:
AQ scores second major win, this time as white

Update 3:

Third game I gave Leela three handicap stones, AQ lost via resignation.