Shoreline wrote:Just report to all of you guys that Facebook dark forest go is a disappointment for me. I expect it to be a strong amateur dan level AI, but instead,
I think it is quite weak.
"The distributed version, named darkfmcts3 in KGS Go Server, use darkfores2 as the underlying DCNN model, runs 75000 rollouts on 2048 threads and produces a move every 13 seconds with one Intel Xeon E5-2680 v2 at 2.80GHz and 44 NVidia K40m GPUs. It uses top-3 predictions in the first 140 moves and switched to top-5 afterwards so that MCTS could have more choices. Pondering is used. Dynamic komi is used only for high handicap games (≥H5). darkfmcts3 now holds a stable KGS 5d level."
I am not sure what "2048 threads" means because Xeon E5-2680v2 is a 16 thread machine. Maybe they meant 128 cores? That would make more sense as I don't see how you can fit 44 GPUs in one machine.
If that is true, it means it is 5 dan with 128 cores and 44 GPUs. So it is not surprising that with your hardware it is very weak. Maybe AlphaGo would be disappointing as well.
A CPU has threads so 2048 indicates a large number of CPU's. A single CPU can have multiple cores and a core can have multiple threads (by way of Intels Hyperthreading). A system can also have multiple CPU's.
A thread can be seen as a train of thought, having multiple threads means you can think of as many things at the same time.
2048 threads is most likely a cluster of servers though, used by way of distributed computing where the workload is spread among multiple systems for increased efficiency vs lower cost (10 servers with cumulative specs are cheaper than 1 with the same sum of specs).
Shoreline wrote:Just report to all of you guys that Facebook dark forest go is a disappointment for me. I expect it to be a strong amateur dan level AI, but instead,
I think it is quite weak.
"The distributed version, named darkfmcts3 in KGS Go Server, use darkfores2 as the underlying DCNN model, runs 75000 rollouts on 2048 threads and produces a move every 13 seconds with one Intel Xeon E5-2680 v2 at 2.80GHz and 44 NVidia K40m GPUs. It uses top-3 predictions in the first 140 moves and switched to top-5 afterwards so that MCTS could have more choices. Pondering is used. Dynamic komi is used only for high handicap games (≥H5). darkfmcts3 now holds a stable KGS 5d level."
I am not sure what "2048 threads" means because Xeon E5-2680v2 is a 16 thread machine. Maybe they meant 128 cores? That would make more sense as I don't see how you can fit 44 GPUs in one machine.
If that is true, it means it is 5 dan with 128 cores and 44 GPUs. So it is not surprising that with your hardware it is very weak. Maybe AlphaGo would be disappointing as well.
A CPU has threads so 2048 indicates a large number of CPU's. A single CPU can have multiple cores and a core can have multiple threads (by way of Intels Hyperthreading). A system can also have multiple CPU's.
A thread can be seen as a train of thought, having multiple threads means you can think of as many things at the same time.
2048 threads is most likely a cluster of servers though, used by way of distributed computing where the workload is spread among multiple systems for increased efficiency vs lower cost (10 servers with cumulative specs are cheaper than 1 with the same sum of specs).
I did not pay attention that Facebook dark forest go requires that much computer resource. If darkforstgo only achieve KGS 5d with 2048 threads, that is not impressive at all. I think alphago would do much better in that setting.
I should take back what I said before on Facebook AI. Facebook had recent bug fix and obviously the strength has been restored. My new testing
on 5000 rollout per move setting put DarkforestGo at some where around DGS 3 Dan level. This would put DarkforestGo at 3rd strongest open source
go AI behind RN3.9 and NG06.