[Computer-go] intel i5 760 vs amd Phenom II X6 1055T
pasky at ucw.cz
Mon Nov 29 05:08:50 PST 2010
On Mon, Nov 29, 2010 at 08:02:10AM -0500, steve uurtamo wrote:
> > OTOH the tree search part is entirely different, very heavy on floating
> > point arithmetics (IIRC, most of the time in this part is spent taking
> > the square roots; overally, tree search is taking 1/4 to 1/3 of the
> > runtime, a significant amount; of course the balances may be different
> > in other programs with heavier/lighter playouts).
> are these float operations necessary?
You have to get the result properly. At least I found out that if I
start cheating on the UCB math, the performance plummets.
> if they're taking place at the
> same time as independent integer operations, pipelining on some cpus
> could take advantage of that,
No, either you walk the tree or you are doing the simulations. :)
Besides, AFAIK most CPUs can do just a single FP op at once?
> but otherwise it seems like a lookup table would be faster.
In Pachi, I found out that (at least on the particular machine I
benchmarked it at back then) lookup table for integer x<32 works best
but is not worth it above that. But you still have to do many divisions
which are also slow.
Maybe I was too hasty with the statement that sqrt is what takes the
longest, it's more than a year since I benchmarked this the last time
and I probably don't remember it properly anymore.
Petr "Pasky" Baudis
The true meaning of life is to plant a tree under whose shade
you will never sit.
More information about the Computer-go