I've never heard of this before? Pretty sure machines are sent to reviewers not indexed, which would explain why I've consistently seen scores rise over time compared to early reviews generally speaking.
Funny, I would say the opposite, I've found that some of the early review sites have the highest scores, while its user-scores that sometimes rise over time - with variation as new machines trickle out into the wild. I can't remember where I saw it, it was some video review where they mentioned either they get machines with very short indexing or the indexing already done. This was years ago though. So maybe I'm wrong or maybe Apple has changed policies. I would be surprised since Apple would basically be kneecapping themselves by not giving reviewers enough time to run benchmark suites without background tasks mucking up the results. Regardless, if this were something Apple has always done previous reviews of previous products would show the same issues ... and they don't ... and it was not just one reviewer. Don't jump to blaming reviewers if multiple ones that are highly experienced are all reporting variability that appears out of the ordinary or showing scores that don't see the expected improvement.
Also I think M5 Pro basically matching Intel Core 285K on Cinebench is impressive, because it's using 6 less cores total (and Cinebench is a rudimentary test that just assumes the more cores you throw at a task the better it is), and on top of that is using 2 less high performance cores (6/12 vs 8/16 for intel).
So on Geekbench 6 it besting it out on more authentic multi-core workloads at 28-29K vs 22K when it's using 6 less cores and 2 less HP cores is impressive to me. And does so with a lot less watts
That's not a fair assessment of either GB or CB. Cinebench is not more or less rudimentary than GB. It tests sustained performance in one particular workload - 3D rendering - and does so in a particular stress-heavy way (if you run the default test). It's basically the equivalent of Blender but using the Redshift engine instead of Blender's cycles engine (Blender also does three scenes to CB's one looped over). Thus on the GPU and CPU it tests heavy FP and vector-processing workloads that can scale very well with cores (though I hypothesize one difference between CB 24 and 26 is the latter does so better). That's also why 3D rendering is a good fit for the GPU and it is also decently representative of a lot of workstation-style FP tasks. But it's only testing that. On CB it is also possible to control the number of threads on which the test runs in order to test how differently threaded applications run. It's just that most people only do ST and full MT (and I suppose now in CB R26 "SC" for devices with SMT2).
GB measures burst rather than sustained performance over many different subtests, including btw 3D rendering which in GB 6 scales across cores/threads just as well as in Blender or CB. In a change from GB5, for GB 6, it's true that not all the MT subtests scale with cores/threads - some do like 3D rendering and compilation, some don't. This was a conscious choice by Primate Labs as a number of consumer applications don't scale with cores/threads. Thus having a large number of cores/threads is not necessarily advantage in this test and it's why a smaller number of more powerful cores can outcompete a larger set of weaker cores which makes Apple's dominance of larger CPUs in MT the expected behavior - even with Apple's new SOC design. However, this is moderated somewhat by the fact that x86 machines often have a greater discrepancy between burst and sustained performance which allows x86 to use those huge spikes in ST and MT clocks to close the distance. Further, the top-line ST and MT numbers are geometric means of the subtests but themselves are not hugely meaningful - same for SPEC btw which GB loosely bases itself on (though SPEC still splits FP and INT workloads, GB does a weighted average of the two). We all use the top line numbers as a shorthand, but if you really want to get into the performance comparisons, then it's the subtest scores that really matter - some of which again scale with cores, some do not. Unfortunately reporting all the subtests is much more unwieldy, which is why people often just report the geometric means. Now this doesn't mean that GB is less authentic either, whenever I see someone complaining about GB and calling it "synthetic" (it is not), I explain its design as above and point out why Primate Labs did what they did when they switched to GB 6.
In summary, CB tests sustained performance on one type of workload, while GB measures burst performance on many different workloads. Each approach has strengths and weaknesses, both are fine to measure and use in the appropriate context. Neither should be taken as gospel.
The only CPU test I can actually think of that is still in common use, and I don't know why it is so, that qualifies as rudimentary or even "synthetic" is Passmark.
Especially the subtests. When @leman and I did violin plots of the subtests, the outliers were horrible both in number and extent - those charts are floating around these forums somewhere. One expects some differences between machines, silicon lottery and all that, but GB is very sensitive to it. The median is fine, but I suspect the very short runtimes is the problem there. But yes, some of this early stuff is going to be from people benchmarking a machine that is still indexing. I still use GB, but yeah ... I'd prefer slightly longer tests to cut down the noise, but maybe their general audience wouldn't.
