for kai laskos (about doubling speed elo gain)

[stavros]

i know that for every doubling speed the elo gain is around 50 elo at blitz level
i dont remember, you had some data in the past, this elo can change according time control
my question is: for long time control for example 90 min/game 120 mins/40 moves and so on..
can you estimate the elo gain at that time control?

(ofcourse anyone with some data can reply)

[jp]

Yeah, it'd be great if there were something better than that rule of thumb that let the user match (by time/node handicap) two engines' playing strengths as closely as possible.

[pohl4711]

http://www.fastgm.de/time-control4.html

[Laskos]

i know that for every doubling speed the elo gain is around 50 elo at blitz level
i dont remember, you had some data in the past, this elo can change according time control
my question is: for long time control for example 90 min/game 120 mins/40 moves and so on..
can you estimate the elo gain at that time control?

(ofcourse anyone with some data can reply)

There is no Elo rule of thumb, there is behavior rule of thumb. Elo gain from doubling decreases with time control times hardware (with strength).

I think for 1 modern core at CCRL 40/4' blitz TC, top engines of today would show some 90 Elo points gain from doubling.
1 modern core CCRL 40/40' TC, top engines would show some 60 Elo points per doubling.
4 modern cores 40/40' TC maybe 50 Elo points.
16 modern cores 40/120' LTC maybe 30 Elo points.
In current TCEC conditions some 25 Elo points from doubling.

On the opposite, for very fast time control, the doubling can reach as much as 200 Elo points at 10s + 0.1s TC on one core.

This is not that hard. The harder thing seems to be to get the behavior of Win/Loss ratio to LTC. From the latest months' data of Andreas, I seem to get that Win/Loss ratio increases from rapid to LTC, almost consistently already. It might be (not very likely) related to the contempt used recently by many top engines.

[jp]

Some other data Kai produced previously for node tripling ---

Lc0 at 3k nodes/move versus Lc0 at 1k nodes/move:

Score of lc0_32930_3k vs lc0_32930_1k: +53 -3 =44 [0.750] 100
Elo difference: 190.85 +/- 51.12
Finished match

SF_dev at 3 million nodes/move versus SF_dev at 1 million nodes/move:

Score of SF_dev_3M vs SF_dev_1M: +60 -4 =36 [0.780] 100
Elo difference: 219.87 +/- 57.61
Finished match

That's much more than 50 elo per doubling!

[jp]

I think for 1 modern core at CCRL 40/4' blitz TC, top engines of today would show some 90 Elo points gain from doubling.
1 modern core CCRL 40/40' TC, top engines would show some 60 Elo points per doubling.
4 modern cores 40/40' TC maybe 50 Elo points.
16 modern cores 40/120' LTC maybe 30 Elo points.
In current TCEC conditions some 25 Elo points from doubling.

Roughly how many nodes per move would these CCRL conditions work out to?
They say the benchmark for their TC is AMD X2 4600+ at 2.4GHz.

TCEC hardware looks like about 10-100 Mnps.

[Laskos]

I think for 1 modern core at CCRL 40/4' blitz TC, top engines of today would show some 90 Elo points gain from doubling.
1 modern core CCRL 40/40' TC, top engines would show some 60 Elo points per doubling.
4 modern cores 40/40' TC maybe 50 Elo points.
16 modern cores 40/120' LTC maybe 30 Elo points.
In current TCEC conditions some 25 Elo points from doubling.

Roughly how many nodes per move would these CCRL conditions work out to?
They say the benchmark for their TC is AMD X2 4600+ at 2.4GHz.

TCEC hardware looks like about 10-100 Mnps.

The CCRL conditions are just time control, NPS is higher on one modern core than on one CCRL core, say 1.5-1.8 MNPS on one core for Stockfish.
The numbers I presented are pretty rough and are valid for top modern AB engines.

[jp]

http://www.fastgm.de/time-control4.html

Thanks for the link. In the graph (Komodo 9.3) the losses are pretty constant. The win decrease with longer times roughly equals the draw increase.

[jp]

... which is relevant to the W/L question Kai raised (though it's only Komodo 9.3).

This is not that hard. The harder thing seems to be to get the behavior of Win/Loss ratio to LTC. From the latest months' data of Andreas, I seem to get that Win/Loss ratio increases from rapid to LTC, almost consistently already. It might be (not very likely) related to the contempt used recently by many top engines.

[Laskos]

... which is relevant to the W/L question Kai raised (though it's only Komodo 9.3).

This is not that hard. The harder thing seems to be to get the behavior of Win/Loss ratio to LTC. From the latest months' data of Andreas, I seem to get that Win/Loss ratio increases from rapid to LTC, almost consistently already. It might be (not very likely) related to the contempt used recently by many top engines.

Yes, but the data is contradictory when the whole FGRL data are used. Also, I am getting on my hardware higher W/L ratio for Lc0 T40 against SF at 4m + 4s compared to 1m + 1s TC from balanced openings, but pretty stable Elo difference. Not LTC at all, but the issue is not easy to me. Also, self-play might have a bit different behavior compared to the general case. Contempt too might interfere, as it's lowering the draw rate, but to unclear proportion of wins and losses.