Has anyone been able to get it to run on a 1080 series gpu? I know it's not officially supported but I was hoping maybe it would run.
Thus far it won't even move a piece for me.
ceres
Moderators: hgm, Rebel, chrisw
-
- Posts: 1142
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: ceres
I think TCEC games is a horrible way to rate the engine's strength, specially the SUFI engines. The sample is too small and the openings they play are ridiculous to meassure strength (they are great for spectacle/show though). I believe with more "normal" openings, as seen on DivP, the difference "IN TCEC CONFIG/HARDWARE" is much less than that.
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls
-
- Posts: 138
- Joined: Thu Jun 18, 2020 9:20 pm
- Full name: Finn Eggers
Re: ceres
Also the discussion about what is fair and what isnt really annoys me. Like how could one even start to compare gpu to cpu.amanjpro wrote: ↑Wed Aug 11, 2021 9:55 pmHuh? did I say you were wrong in the estimate? I said the diff between current Leela and SF in TCEC rating is about 20 points anyways, so +20 for Ceres means something at this level....Luecx wrote: ↑Wed Aug 11, 2021 9:50 pmHow about you start asking the actual authors instead of unknowing people in this forum? I asked in the Leela Discord a week ago. 20 ELO. I assume the Leela guys know how to properly test gpu engines.amanjpro wrote: ↑Wed Aug 11, 2021 7:05 pmDo we have a rating list that is fair to GPU engines? according to TCEC, SF is around 20 points stronger than Leela... all other rating lists, that I know off, test CPU version of Leela, which is not near its top top strength really...
It seems to me as if there are two aproaches and I do not know which of them is the better one.
Either we compare CPU and GPU engines based on OPS. It seems to make total sense, right? Looking at which engine performs better given the same amount of operations per second. Sadly this will look very very very bad for GPU engines since basically any GPU outperforms the strongest CPUs in operations per second.
Secondly, we could compare prices for hardware. Saying that both hardwares should be equal in price. Also makes sense in a way: Which engine is better given the same amount of money. Clearly the race will be a lot closer although I do believe GPU-engines would still lose that battle but I may be wrong.
-
- Posts: 883
- Joined: Sat Mar 13, 2021 1:47 am
- Full name: Amanj Sherwany
Re: ceres
I am not saying we should have a fair rating list, I don't beleive it can be found for GPU and CPU...Luecx wrote: ↑Fri Aug 13, 2021 6:03 pmAlso the discussion about what is fair and what isnt really annoys me. Like how could one even start to compare gpu to cpu.amanjpro wrote: ↑Wed Aug 11, 2021 9:55 pmHuh? did I say you were wrong in the estimate? I said the diff between current Leela and SF in TCEC rating is about 20 points anyways, so +20 for Ceres means something at this level....Luecx wrote: ↑Wed Aug 11, 2021 9:50 pmHow about you start asking the actual authors instead of unknowing people in this forum? I asked in the Leela Discord a week ago. 20 ELO. I assume the Leela guys know how to properly test gpu engines.amanjpro wrote: ↑Wed Aug 11, 2021 7:05 pmDo we have a rating list that is fair to GPU engines? according to TCEC, SF is around 20 points stronger than Leela... all other rating lists, that I know off, test CPU version of Leela, which is not near its top top strength really...
It seems to me as if there are two aproaches and I do not know which of them is the better one.
Either we compare CPU and GPU engines based on OPS. It seems to make total sense, right? Looking at which engine performs better given the same amount of operations per second. Sadly this will look very very very bad for GPU engines since basically any GPU outperforms the strongest CPUs in operations per second.
Secondly, we could compare prices for hardware. Saying that both hardwares should be equal in price. Also makes sense in a way: Which engine is better given the same amount of money. Clearly the race will be a lot closer although I do believe GPU-engines would still lose that battle but I may be wrong.
Leela is clearly the best engine that is not SF... and the distance is not that far from SF, so any improvement on Leela is a big deal... I believe we can't know how close they are to each other in strength, just because as you said there is no way to compare them really
-
- Posts: 1142
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: ceres
There is another approach that might be "fair"... use the same machine for both engines. Then, it's up to the engines to use the resources/parts that they need/want. Because with your criteria, it isn't fair to compare engines that use SMP with engines that don't, or engines that support a huge hash table vs. engines that don't, or engines that support TBs vs engines that don't, or engines that have several features in the search vs engines that don't... where do you draw the line?Luecx wrote: ↑Fri Aug 13, 2021 6:03 pmAlso the discussion about what is fair and what isnt really annoys me. Like how could one even start to compare gpu to cpu.amanjpro wrote: ↑Wed Aug 11, 2021 9:55 pmHuh? did I say you were wrong in the estimate? I said the diff between current Leela and SF in TCEC rating is about 20 points anyways, so +20 for Ceres means something at this level....Luecx wrote: ↑Wed Aug 11, 2021 9:50 pmHow about you start asking the actual authors instead of unknowing people in this forum? I asked in the Leela Discord a week ago. 20 ELO. I assume the Leela guys know how to properly test gpu engines.amanjpro wrote: ↑Wed Aug 11, 2021 7:05 pmDo we have a rating list that is fair to GPU engines? according to TCEC, SF is around 20 points stronger than Leela... all other rating lists, that I know off, test CPU version of Leela, which is not near its top top strength really...
It seems to me as if there are two aproaches and I do not know which of them is the better one.
Either we compare CPU and GPU engines based on OPS. It seems to make total sense, right? Looking at which engine performs better given the same amount of operations per second. Sadly this will look very very very bad for GPU engines since basically any GPU outperforms the strongest CPUs in operations per second.
Secondly, we could compare prices for hardware. Saying that both hardwares should be equal in price. Also makes sense in a way: Which engine is better given the same amount of money. Clearly the race will be a lot closer although I do believe GPU-engines would still lose that battle but I may be wrong.
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls
-
- Posts: 138
- Joined: Thu Jun 18, 2020 9:20 pm
- Full name: Finn Eggers
Re: ceres
Tbh I do not think they are that close. Looking at those insane uneven openings at TCEC and SF still crushing it. I mean many games arent decisive and still SF wins like 54-46 (I do not know the score but i guess it was something like that). I would say there is like a 50 Elo gap but who knowsamanjpro wrote: ↑Fri Aug 13, 2021 6:10 pmI am not saying we should have a fair rating list, I don't beleive it can be found for GPU and CPU...Luecx wrote: ↑Fri Aug 13, 2021 6:03 pmAlso the discussion about what is fair and what isnt really annoys me. Like how could one even start to compare gpu to cpu.amanjpro wrote: ↑Wed Aug 11, 2021 9:55 pmHuh? did I say you were wrong in the estimate? I said the diff between current Leela and SF in TCEC rating is about 20 points anyways, so +20 for Ceres means something at this level....Luecx wrote: ↑Wed Aug 11, 2021 9:50 pmHow about you start asking the actual authors instead of unknowing people in this forum? I asked in the Leela Discord a week ago. 20 ELO. I assume the Leela guys know how to properly test gpu engines.amanjpro wrote: ↑Wed Aug 11, 2021 7:05 pmDo we have a rating list that is fair to GPU engines? according to TCEC, SF is around 20 points stronger than Leela... all other rating lists, that I know off, test CPU version of Leela, which is not near its top top strength really...
It seems to me as if there are two aproaches and I do not know which of them is the better one.
Either we compare CPU and GPU engines based on OPS. It seems to make total sense, right? Looking at which engine performs better given the same amount of operations per second. Sadly this will look very very very bad for GPU engines since basically any GPU outperforms the strongest CPUs in operations per second.
Secondly, we could compare prices for hardware. Saying that both hardwares should be equal in price. Also makes sense in a way: Which engine is better given the same amount of money. Clearly the race will be a lot closer although I do believe GPU-engines would still lose that battle but I may be wrong.
Leela is clearly the best engine that is not SF... and the distance is not that far from SF, so any improvement on Leela is a big deal... I believe we can't know how close they are to each other in strength, just because as you said there is no way to compare them really
-
- Posts: 138
- Joined: Thu Jun 18, 2020 9:20 pm
- Full name: Finn Eggers
Re: ceres
I do think that is actually fundamentally unfair in the fact that machines badly differ in their strength. Like a machine has lots of parts which the engine may use or may not. I am talking about the motherboard, ram, avx, gpu, cpu, ..... The list goes on and seems endless. The engine strength would purely rely on the characteristic of a single machine.CMCanavessi wrote: ↑Fri Aug 13, 2021 6:56 pmThere is another approach that might be "fair"... use the same machine for both engines. Then, it's up to the engines to use the resources/parts that they need/want. Because with your criteria, it isn't fair to compare engines that use SMP with engines that don't, or engines that support a huge hash table vs. engines that don't, or engines that support TBs vs engines that don't, or engines that have several features in the search vs engines that don't... where do you draw the line?Luecx wrote: ↑Fri Aug 13, 2021 6:03 pmAlso the discussion about what is fair and what isnt really annoys me. Like how could one even start to compare gpu to cpu.amanjpro wrote: ↑Wed Aug 11, 2021 9:55 pmHuh? did I say you were wrong in the estimate? I said the diff between current Leela and SF in TCEC rating is about 20 points anyways, so +20 for Ceres means something at this level....Luecx wrote: ↑Wed Aug 11, 2021 9:50 pmHow about you start asking the actual authors instead of unknowing people in this forum? I asked in the Leela Discord a week ago. 20 ELO. I assume the Leela guys know how to properly test gpu engines.amanjpro wrote: ↑Wed Aug 11, 2021 7:05 pmDo we have a rating list that is fair to GPU engines? according to TCEC, SF is around 20 points stronger than Leela... all other rating lists, that I know off, test CPU version of Leela, which is not near its top top strength really...
It seems to me as if there are two aproaches and I do not know which of them is the better one.
Either we compare CPU and GPU engines based on OPS. It seems to make total sense, right? Looking at which engine performs better given the same amount of operations per second. Sadly this will look very very very bad for GPU engines since basically any GPU outperforms the strongest CPUs in operations per second.
Secondly, we could compare prices for hardware. Saying that both hardwares should be equal in price. Also makes sense in a way: Which engine is better given the same amount of money. Clearly the race will be a lot closer although I do believe GPU-engines would still lose that battle but I may be wrong.
This would only be fair, if and only if that one machine would be specified in the first place and everyone would optimise for it. This would not lead to gpu and cpu engines coexisting.
-
- Posts: 1142
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: ceres
It always does. Doesn't matter how you look at it.
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls
-
- Posts: 138
- Joined: Thu Jun 18, 2020 9:20 pm
- Full name: Finn Eggers
Re: ceres
the only way to properly test is having 2 machines like what tcec does for gpu vs cpu where the gpu machine is optimised for gpu engines and vice versa. Here the question arises (what i brought up to begin with), how both hardwares should be "equallized". same price, ops, something else?
-
- Posts: 1142
- Joined: Thu Dec 28, 2017 4:06 pm
- Location: Argentina
Re: ceres
Pick whatever criteria you want. There's not one inherently better than the other. You can also use electrical power consumed, etc.Luecx wrote: ↑Fri Aug 13, 2021 9:26 pmthe only way to properly test is having 2 machines like what tcec does for gpu vs cpu where the gpu machine is optimised for gpu engines and vice versa. Here the question arises (what i brought up to begin with), how both hardwares should be "equallized". same price, ops, something else?
Follow my tournament and some Leela gauntlets live at http://twitch.tv/ccls