TCEC Division 3 results simulator

[Werewolf]

In this particular case GV100 (we don't know which) is roughly around 8x compared to my OC'ed 1060

Surely more like 3x. Unless it's using the tensor cores (which I thought haven't been developed yet for Lc0) ?

[Milos]

In this particular case GV100 (we don't know which) is roughly around 8x compared to my OC'ed 1060

Surely more like 3x. Unless it's using the tensor cores (which I thought haven't been developed yet for Lc0) ?

It's much more due to half-precision ops (GF16). Lc0 has support for GF16 operations in kernels.
GF16 is twice faster than normal single-precision ops (GF32). That's the reason GV100 is more than twice faster than 1080Ti (actually around 3x). And another 2.5x speed up is difference between 1060 and 1080Ti.

[Werewolf]

In this particular case GV100 (we don't know which) is roughly around 8x compared to my OC'ed 1060

Surely more like 3x. Unless it's using the tensor cores (which I thought haven't been developed yet for Lc0) ?

It's much more due to half-precision ops (GF16). Lc0 has support for GF16 operations in kernels.
GF16 is twice faster than normal single-precision ops (GF32). That's the reason GV100 is more than twice faster than 1080Ti (actually around 3x). And another 2.5x speed up is difference between 1060 and 1080Ti.

I didn't know about the Lc0 support for half precision. Presumably this means a Titan will also be much faster than a 1080 Ti etc?

[chrisw]

Well, LC0 finally managed to get a result, slightly decreasing its draw rate to "only" 78% of all games.

Meanwhile Ethereal dead cert at 100%

And three engines in with a chance of second place and promotion
DeusX 47%
Arasan 27%
Pedone 20%

And LC0 looking sorry for itself at 4%. A couple of LC0 wins, or even a win and draw would put LC0 back in some sort of contention. It next plays Arasan and then DeusX.

[Laskos]

Well, LC0 finally managed to get a result, slightly decreasing its draw rate to "only" 78% of all games.

Meanwhile Ethereal dead cert at 100%

And three engines in with a chance of second place and promotion
DeusX 47%
Arasan 27%
Pedone 20%

And LC0 looking sorry for itself at 4%. A couple of LC0 wins, or even a win and draw would put LC0 back in some sort of contention. It next plays Arasan and then DeusX.

Again seems a bit extreme. I would feel like Deus has 40% (close to your result, but a bit smaller), as it has good SB. But Lc0 still some 10% (much larger than your 4%). Lc0 for its weak score, has good SB, so if it manages to win some, its SB will be fine.

[chrisw]

I didn't test how often SB comes into play for TCEC tie breaks, but it is quite a way down the list, which is:

score
disconnect count
wins
black wins
sb

it would be a bit of hassle to find the rate of sb tiebreak decisions, I could try later maybe, but it is probably quite low. LC0 will have tie break problems because of the low win count. Right now, apart from the bottom two, LC0 is effectively another 0.5 pts behind the rest of the pack, down to this effect.

[Laskos]

I didn't test how often SB comes into play for TCEC tie breaks, but it is quite a way down the list, which is:

score
disconnect count
wins
black wins
sb

it would be a bit of hassle to find the rate of sb tiebreak decisions, I could try later maybe, but it is probably quite low. LC0 will have tie break problems because of the low win count. Right now, apart from the bottom two, LC0 is effectively another 0.5 pts behind the rest of the pack, down to this effect.

Ah, so SB are down the list? Didn't know. Anyway, some strange rules, with a disconnect you not only lose the game, but are also penalized on the first in the list tie-breaks. Pretty severe double-punishment.

[jkiliani]

I didn't test how often SB comes into play for TCEC tie breaks, but it is quite a way down the list, which is:

score
disconnect count
wins
black wins
sb

it would be a bit of hassle to find the rate of sb tiebreak decisions, I could try later maybe, but it is probably quite low. LC0 will have tie break problems because of the low win count. Right now, apart from the bottom two, LC0 is effectively another 0.5 pts behind the rest of the pack, down to this effect.

I think you forgot direct encounter in here. Stated at http://www.chessdom.com/tcec-season-13- ... f-the-nns/, the order is:

score
disconnect count
direct encounter
wins
black wins
Sonneborn-Berger
discretion of tournament director

[chrisw]

I didn't test how often SB comes into play for TCEC tie breaks, but it is quite a way down the list, which is:

score
disconnect count
wins
black wins
sb

it would be a bit of hassle to find the rate of sb tiebreak decisions, I could try later maybe, but it is probably quite low. LC0 will have tie break problems because of the low win count. Right now, apart from the bottom two, LC0 is effectively another 0.5 pts behind the rest of the pack, down to this effect.

Ah, so SB are down the list? Didn't know. Anyway, some strange rules, with a disconnect you not only lose the game, but are also penalized on the first in the list tie-breaks. Pretty severe double-punishment.

Yes, Nemorino is down an effective 0.5 pts at tie breaks because of the disconnect, plus another 1 pt down because the disconnect game was winning. So we must take care not to discount Nemorino.

[chrisw]

I didn't test how often SB comes into play for TCEC tie breaks, but it is quite a way down the list, which is:

score
disconnect count
wins
black wins
sb

it would be a bit of hassle to find the rate of sb tiebreak decisions, I could try later maybe, but it is probably quite low. LC0 will have tie break problems because of the low win count. Right now, apart from the bottom two, LC0 is effectively another 0.5 pts behind the rest of the pack, down to this effect.

I think you forgot direct encounter in here. Stated at http://www.chessdom.com/tcec-season-13- ... f-the-nns/, the order is:

score
disconnect count
direct encounter
wins
black wins
Sonneborn-Berger
discretion of tournament director

wooo!! I missed that. Thanks for the spot. Will fix.