kranium wrote:cool, nice work
thx Andreas
it's nice to see some real data on this subject
is it fair to conclude:
1) extra threads help more in short TCs than in LTC?
2) as threads increase, it appears the graph should flatten out eventually...
i.e. at one point, adding extra cores/threads simply won't increase ELO in any meaningful way (law of diminishing returns)
1) Depends how you define "help". Maybe they help the same way, in terms of average time to depth. But we are simply looking at the normal diminishing returns, in elo terms, of increasing NPS on a single core test. So an effect completely unrelated to SMP. We need to measure average TTD to answer that question, not ELO.
2) Certainly. It's the compounding effect of (at least) two well known effects:
* SMP performance (expressed in terms of in 1/TTD=f(threads)) is less than linear with the number of threads. In theor 1/TTD should follow an Ahmdal's law. I remember Miguel made some pretty good fit with that.
* Even on a single threaded program, increasing NPS is less than linear in ELO terms. First time you double NPS, maybe you get +100 elo, second time, maybe only 90 elo etc.
And, as always, long TC flattens everything, because draw rate increases. This is inevitable. In a theoretical sense, when tc --> infinity, you get perfect play, hence 100% draws and 0 ELO difference.
Theory and practice sometimes clash. And when that happens, theory loses. Every single time.