For the most part, I believe CCRL currently uses actual time controls closer to 40/20 rather than 40/40 due to the huge advancements made in CPU hardware.
A modern quad core at 40/20 would be comparable to a 1050 at 40/20 because even though both pieces of hardware are much newer than the Athlon 4600+, they both have half as much time to match.
The problem of different GPU models could potentially be solved with asymmetric time controls.
e.g. CCRL 40/40 could mean:
40 moves in 40 minutes with an Athlon64 4600+, benchmarked with (some named) version of Crafty
40 moves in 20 minutes with a GTX 1050, benchmarked with (some named) version of Lc0
konsolas wrote: ↑Mon Feb 25, 2019 11:02 pm
For the most part, I believe CCRL currently uses actual time controls closer to 40/20 rather than 40/40 due to the huge advancements made in CPU hardware.
A modern quad core at 40/20 would be comparable to a 1050 at 40/20 because even though both pieces of hardware are much newer than the Athlon 4600+, they both have half as much time to match.
The problem of different GPU models could potentially be solved with asymmetric time controls.
e.g. CCRL 40/40 could mean:
40 moves in 40 minutes with an Athlon64 4600+, benchmarked with (some named) version of Crafty
40 moves in 20 minutes with a GTX 1050, benchmarked with (some named) version of Lc0
Ok, but how do you deal with the reality that GPU based engines also need a CPU? While Leela is certainly quite GPU heavy, I am quite sure even in this case the CPU is non-negligible. To make matters worse, you could easily imagine an engine which is roughly evenly bottlenecked by the GPU and CPU. How are you going to deal with that?
The GTX 1050 has been bought because of low cost and power efficiency and with decent performance, and that my 8 year old noisy and power hungry GPU needs replacing.
However it then opens up the opportunity to try some blitz-only GPU testing (not 40/40). In the short term it is easy enough but yes there are medium to longer term issues to resolve as well. I think it is a good fit for the list, but if it isn't welcomed then I simply won't bother.
jorose wrote: ↑Mon Feb 25, 2019 11:32 pmTo make matters worse, you could easily imagine an engine which is roughly evenly bottlenecked by the GPU and CPU. How are you going to deal with that?
You could benchmark with the individual engine, and give each GPU-using engine its own time-odds factor depending on its speed on the standard and actual hardware.
konsolas wrote: ↑Mon Feb 25, 2019 11:02 pm
For the most part, I believe CCRL currently uses actual time controls closer to 40/20 rather than 40/40 due to the huge advancements made in CPU hardware.
A modern quad core at 40/20 would be comparable to a 1050 at 40/20 because even though both pieces of hardware are much newer than the Athlon 4600+, they both have half as much time to match.
The problem of different GPU models could potentially be solved with asymmetric time controls.
e.g. CCRL 40/40 could mean:
40 moves in 40 minutes with an Athlon64 4600+, benchmarked with (some named) version of Crafty
40 moves in 20 minutes with a GTX 1050, benchmarked with (some named) version of Lc0
CCRL is 40/20 with my i5-4460. (I benchmarked it)
I think it's closer to 40/15 or 40/12 with a current gen i5-9600.
Crafty 19.17 BH benchmark can be downloaded here. (Version 19.17, Brian Hoffman compile, 32-bit, single-CPU). Please note that we should use the same version and compile because different versions may be slightly faster or slower and will give different benchmark time.
but the link is a windows only version. I guess some CCRL members are testing under linux ?
