Finally, note how several systems show little or no negative deviation in the areas were faster and hotter systems fall precipitously. In one area, the 1.26GHz P4 outperforms the 1.9Ghz P4 by 31%.
Is this conclusive proof of hot spot throttling on the Matlap Linpack test? No, and to a typical user it doesn’t matter if throttling caused these abnormal performance curves or not. The most important fact is simply that the Pentium 4 displays, yet again, quirky performance that can often fall far short of expectations. Furthermore these deviations often do not seem to make sense based on common performance evaluation criteria.
In fact, our P4 Linpack results demonstrate one reason why the P4 has such erratic behavior on different tasks. For instance, a Matlab Linpack application designed to operate on a 87kB array might actually see substantially worse performance on a 1.9Ghz P4 than on a 1. 4GHz P4. For anyone using this package combination on the P4, beware that this chip’s performance curve is chock full of potholes. These results were reproducible on all CPUs tested without respect to ambient temperature, heat sink quality, motherboard type, etc.
inqst.com
Fascinating set of graphs accompanies the article.
P4 may be OK once it's on .13, or maybe it won't be. But it doesn't appear to be a reliable performer on anything but NOPs when it's fabbed on .18. Those few times software actually needs a P4 at anything over about 1.26GHZ, the chip will often have to throttle back or burn up. It may not make much of a real world difference for most people, but it's pretty clear that there's no point in paying for a P4 clocked any higher than 1.3GHZ - they don't run any faster than that under load, anyway.
Kudos to Scumbria for picking up that the double clocked units were just asking for trouble. |
