Page 1 of 1

ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Thu Jul 11, 2019 11:42 pm
by Hai
INT8 is much better than FP16.

Some days ago at TCEC:
Scorpio INT8 vs LCZero1pct v0.21.2-n42547 +0 =4 -0
Scorpio INT8 vs LCZeroCPU v0.21.2-nLD2 +0 =4 -0

Now Scorpio INT8 vs LCZeroCPU v0.21.2-nLD2 +1 =1 -0

Will LC0 ever use INT8 instead of FP16?

Re: ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Fri Jul 12, 2019 12:02 am
by Dann Corbit
I don't think we can conclude superiority from ten games.

Logically, though, faster is better. I guess that if Scorpio starts drilling LC0 into the ground, they will look at what Scorpio is doing and try to imitate it.

Re: ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Fri Jul 12, 2019 6:45 am
by Chessqueen
Hai wrote: Thu Jul 11, 2019 11:42 pm INT8 is much better than FP16.

Some days ago at TCEC:
Scorpio INT8 vs LCZero1pct v0.21.2-n42547 +0 =4 -0
Scorpio INT8 vs LCZero1pct +0 =4 -0

Now Scorpio INT8 vs LCZeroCPU v0.21.2-nLD2 +1 =1 -0

Will LC0 ever use INT8 instead of FP16?

You are aware that LCZero1pct v0.21.2-n42547 was only using 1 pct or percent, LOOK CLOSELY.... LCZero1pct, and...... in this case LCZeroCPU v0.21.2-nLD2 +1 =1 -0 was using the cpu with a smaller Net, NOT the GPU version, therefore, the cpu version performs much worse at least 400 Elo lower than the GPU version at 100% percent :shock: :mrgreen:

Re: ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Fri Jul 12, 2019 8:14 am
by Uri Blass
Chessqueen wrote: Fri Jul 12, 2019 6:45 am
Hai wrote: Thu Jul 11, 2019 11:42 pm INT8 is much better than FP16.

Some days ago at TCEC:
Scorpio INT8 vs LCZero1pct v0.21.2-n42547 +0 =4 -0
Scorpio INT8 vs LCZero1pct +0 =4 -0

Now Scorpio INT8 vs LCZeroCPU v0.21.2-nLD2 +1 =1 -0

Will LC0 ever use INT8 instead of FP16?

You are aware that LCZero1pct v0.21.2-n42547 was only using 1 pct or percent, LOOK CLOSELY.... LCZero1pct, and...... in this case LCZeroCPU v0.21.2-nLD2 +1 =1 -0 was using the cpu with a smaller Net, NOT the GPU version, therefore, the cpu version performs much worse at least 400 Elo lower than the GPU version at 100% percent :shock: :mrgreen:
Ethereal 11.50 lost 2.5-1.5 against LCZeroCPU v0.21.2-nLD2 so I guess that the difference between the cpu version and the gpu version is less than 400 elo.

Edit:Xiphos0.5.3 also drew 2-2 against the same cpu version and Xiphos0.5 is number 5 in the CCRL 40/40 rating list

Re: ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Fri Jul 12, 2019 4:35 pm
by dkappe
ScorpioNN is good and might make it to DivP. The weakest part is probably the NN. It needed another 10-20 million self-play games to catch up to t30, but I wasn’t willing to shell out on vast.ai resources to get there. So until someone cooks up a better net, it’s likely to be a bit weaker than lc0/Allie/etc.

Also, Scorpio uses custom TB’s which are not present on TCEC, so it is essentially playing without them.

Re: ScorpioNN v2.9.0-n_maddex_INT8 is great:)

Posted: Fri Jul 12, 2019 10:49 pm
by supersharp77
dkappe wrote: Fri Jul 12, 2019 4:35 pm ScorpioNN is good and might make it to DivP. The weakest part is probably the NN. It needed another 10-20 million self-play games to catch up to t30, but I wasn’t willing to shell out on vast.ai resources to get there. So until someone cooks up a better net, it’s likely to be a bit weaker than lc0/Allie/etc.

Also, Scorpio uses custom TB’s which are not present on TCEC, so it is essentially playing without them.
Scorpio NN settings are very complex.ditto for Nebiyu chess.....having trouble getting the NN's to interface properly with the engines...actually LC0 is easier to set up these days! Think a Scorpio NN autodetect feature is called for..:) :wink: