Ali: And exactly what the current formula is?
Model rating = 1300 + (f - 1200)*100/66
Each "successive addition" changes the CPU core frequency in less and less diminished steps (in %%-wise); therefore the difference between consecutive scores should not drop faster than it started at 1500+, which was 0.1/step
I would argue that you are wrong… At very low frequencies, the CPU is (almost) never waiting for anything external and scaling is therefore completely linear. That is, the drop in "difference between consecutive scores" (d²P/df²) is zero.
At higher frequencies, the probability of the CPU having to wait for accesses to main memory increases and becomes significant. At this point d²P/df² > 0.
This contradicts your claim that "the difference
between consecutive scores should not drop faster than
it started at".
Remember, this has nothing to do with quantispeed. The increase in (real) frequency between successive model numbers is a constant 66MHz.
no matter how funny it sounds, but the ability
of P4 to "scale better" is a direct consequence of it's
lower IPC.
…
In this case(assuming the memory/FSB still runs at the same frequency), the change in df will make almost no difference in performance, so your dP/df will be zero.
However, in the AthlonXP vs P4 case, there is a huge difference in available bandwidth. This significantly impacts not only performance, but also scaling. Same goes for larger caches, although this is certainly harder to compare directly, since the caches are very different in nature (latency, bandwidth and associativity - along with size).
-fyo |
