Here's some news from the DAB. Please keep this confidential:
Re: Bigadv-16 first beta upcoming
Postby kasson » 10 Feb 2012, 14:42
Ok, based on preliminary data we have a tentative first-round adjustment to the bigadv points. (Notice all the qualifiers here--this is a work in progress.)
Over the bigadv-capable systems we have sampled (reporting 24-128 cores), we're seeing 8101/6904 PPD ratios at about 0.32, 8101/6903 at about 0.33, 8101/6901 at about 0.51, and 6903/6901 at about 1.55. This also confirms what we've known for a while, that 6903 & 6904 were over-valued with respect to 6901 and friends.
WIthout trying to fix everything at once, we're considering the following changes to base points:
8101 -> 2.2x
6903 & 6904 -> 0.8x
6901 & 6900 -> 1.15x
This would take us to 8101/6904 at 0.89, 8101/8903 at 0.91, 8101/6901 at 1.02, 6903/6901 at 1.08.
This is in part correcting for the historical issue of high values for 6903 & 6904, where we reduced points for 6901 as part of an overall bigadv markdown.
Of course, here we're not looking at the slope of the bonus values. We could do a more sophisticated set of fits that consider the k-factors as well, but let's start with this and see how we do...
Re: Bigadv-16 first beta upcoming
Postby kasson » 10 Feb 2012, 14:42
Ok, based on preliminary data we have a tentative first-round adjustment to the bigadv points. (Notice all the qualifiers here--this is a work in progress.)
Over the bigadv-capable systems we have sampled (reporting 24-128 cores), we're seeing 8101/6904 PPD ratios at about 0.32, 8101/6903 at about 0.33, 8101/6901 at about 0.51, and 6903/6901 at about 1.55. This also confirms what we've known for a while, that 6903 & 6904 were over-valued with respect to 6901 and friends.
WIthout trying to fix everything at once, we're considering the following changes to base points:
8101 -> 2.2x
6903 & 6904 -> 0.8x
6901 & 6900 -> 1.15x
This would take us to 8101/6904 at 0.89, 8101/8903 at 0.91, 8101/6901 at 1.02, 6903/6901 at 1.08.
This is in part correcting for the historical issue of high values for 6903 & 6904, where we reduced points for 6901 as part of an overall bigadv markdown.
Of course, here we're not looking at the slope of the bonus values. We could do a more sophisticated set of fits that consider the k-factors as well, but let's start with this and see how we do...