CEP works well on older machines: 106% awarded/claimed for me

In this context, they are extremely meaningful. They are supposed to be even for a given amount of work. Assuming that this is roughly true across the population of WCG machines, different machines clearly perform more efficiently on some types of WUs than others. Therefore, it makes sense to set your machines to do WUs that they are relatively efficient at, not WUs that they suck at.

Very old regurgitated hat, so not going to spend much time on this beyond this post; 64 bit OSses have a double / triple integer benchmark versus 32bit OS equivalent. The flop test is virtually identical to the 32 bit OS, and that´s no surprise for a very good reason.

If 32 bit apps which hardly have integer crunch on 64 bit OSses, they claim high due the integer component of the benchmark, which is just a result of Dhry+Whetstone divvied by about 480. As flops still process at 32 bit, regardless of the 64 bit OS, obviously claims get very often paired down unless meeting up in quorum with another 64 bit machine.

Fair - Not Fair, 100,000 Cobblestones for 1 teraop of calculations is the rule, some machines more efficient at doing them then others, so is live.

Hello,
for completing the remarks about this "seems-to-be" hot topic, I would like to add the following comment.
Despite of claimed / granted credits, it is much more interesting to calculate the granted credit per core per hour.
If the host availability for WCG projects is more or less constant, the granted credit / core / hour should be more or less constant (reproducible).
Indeed the host environmental conditions (temperature) could have an impact on the real computation power. Additionally, the type of project calculation (more integer or more float oriented) combined with the processor properties can have a measurable impact on the average performance.
Some months ago, because one of my hosts was regularly and significantly under granted, I performed some comparisons between projects. Surprisingly, I was not able to identify "THE BEST PROJECT" in terms of earned credits. One project could bring 15% more on a specific host but 10% less on another host. However, it was clearly possible to identify project which brought everywhere significantly less (-20%) credits than the calculated average ! ...
This observation fits with the further observed performance problems (e.g. PF).
If I would have sufficient time (!!!), I would enjoy to investigate this question more deeply because, in order to improve the grid performance, it could be interesting to identify for each project the most appropriate combination of hardware architecture, OS and project ! ... (because I don't play to Euro-Million, I will probably not have enough time in the near future for completing this investigation ).
Cheers,
Yves

I.e. exactly as expected. All the AutoDock projects have a similar value, except HFCC (which you can't have enough data for yet, and even if you do, the credit algorithm relies on the history). HCC is heavily integer based, so the floating point benchmark is irrelevant, and the claimed/granted difference is larger.

So, all as expected, and as discussed many times before. My colleague Jean explained exactly the same thing in this very thread.

And yes, our reply is authoritative.