Moritz (better@not.tell) on 4/14/11 wrote:
---------------------------
>One way of incorporating the memory would be to add a mismatch/starvation factor.

I was thinking of something like:
*(1-1/(( (RAM_BW/Gflops*Regression_value_1)^Regression_value_2) +1))*

this should not conflict too much with the used parameters if the data contains samples with different core/memorylane relations. But you might end up with less meaningful Regression_value_n because the same info will just be smeared over more of them. the error for the samples will go down but the predictive value will not go up.
In the lucky case we would be able to say how much performance would be lost (absolutely) if one would have so much RAM-datarate.
Since the RAM-width is always n*64bit I don't see much flexibility there.