Your points are all valid. It definitely depends on your workload. I use my i7-980x primarily for developing my chess program Komodo, and it constantly is playing automated test games to evaluate program changes. I have a mutl-threaded testing harness that schedules everything nicely and it really screams on this machine. With my application more cores scale extremely well.
However, alpha/beta or MCTS does not scale that well. So if you had a choice of doubling the cores, or doubling the speed of each core, it's not even a close call. But it would probably be about the same for the tester rig. So for this, I would give a lot of weight to the individual speed of each core. Of course the only way to really know is to run the target GO program on different machines to see which one gets the most playouts. But even that is not so simple as it may be that you should tune things differently for different machines. Don On Sun, Nov 7, 2010 at 8:37 PM, Petr Baudis <[email protected]> wrote: > Hi! > > On Sun, Nov 07, 2010 at 07:37:33AM -0500, Don Dailey wrote: > > I think you get the most bang for the buck with the i7-980x, which is > not > > really the same chip as the other Intel chips, and I think they are > coming > > out with improved versions of that if they have not already. This has > 6 > > real cores (and 6 additional pseudo cores. > > Yes, I was looking at i7-970, which is quite similar. (I don't aim at > overclocking too much, I don't think I can fit a lot of cooling into a > 2U rack server.) > > However, now I'm seriously torn between taking one i7-970 or two 6134 > (which should be the same as 6174 just with 8 cores inst. of 12): > > > http://iltsarnews.blogspot.com/2010/03/amds-opteron-6174-12-cores-magny-cours.html > > It seems to bode extremely well in the benchmarks. Two 6134 are just > a bit more expensive than an i7-970, and if I extrapolate the scores > (which is a rther lame, but what can I do?): > > 2x6174 -> 2x6134: 49372*8/12 = 32915 > 2xX5670 -> 1xi7-970: 35868/2 = 17934 (i7-970 has faster clock, > but slower QPI and memory) > > This seems rather clear, even considering the gross inaccuracies this > kind of transformation introduces. So it seems that this boils down to > how bogus the benchmarks above are. ;-) > > Per core / per thread performance is of course a different matter, > but given how very fuzzy and workload-specific this can get and how well > MCTS parallelizes (with shared memory), I think it's more useful to > discuss the total performance. > > Does anyone have datapoints on how good the latest Opterons (i.e. the > 6100 line) really are? (Compared to i7-970 (or similar Xeon) or at least > i7-950.) > > -- > Petr "Pasky" Baudis > The true meaning of life is to plant a tree under whose shade > you will never sit. > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
