Wasn't there a Nuvia white paper or something showing that GeekBench 5 ST/MT scores correlated very strongly (> 90%) with the equivalent SPEC benchmarks on the same CPUs?
Not that SPEC is representative of everyone's workloads, but it's the closest thing there is to a true "industry standard" for cross-platform/architecture performance comparisons (and has been for over 2 decades).
EDIT:
Found it, was even stronger than I remembered! An R-squared of > 0.99 is just plain nuts.
I don't know the extent to which this applies to SPEC 2017, but with the implementation of AVX512 by AMD, GB may no longer be a good cross-platform comparator.
The problem is that GB gives significant weight to the crypto score, and that is significantly accelerated by AVX512. AMD fully implements this, while of course AS doesn't (and I don't think they have anything equivalent yet). With Intel it's more complicated. Someone please correct me if I'm wrong but, IIUC, Intel implemented this is Rocket Lake (gen 11), but then removed it in Alder Lake (perhaps so that those needing it—presumably typically commercial customers—would be forced to buy Xeons, akin to how NVIDIA disabled FP64 on their consumer GPU's).
Anyways, the point is that AVX-512 has little effect on performance for consumers, since most apps consumers use don't implement it. Thus a benchmark that heavily rewards AVX-512 capability (as it appears GB does) might not be reflective of typical consumer workloads.
If the above is correct then, for typical consumer workloads, GB currently overestimates the performance of AMD-7000 and Xeon vs. AS and Intel-Alder Lake.
For more details see:
The scores are real, the real-world performance gains aren't
www.tomshardware.com
And:
www.anandtech.com
From Anandtech:
"For our 3DPM v2.1 testing, we added in the Intel Core i9-11900K (Rocket Lake) to show performance across AVX workloads. Although
Intel officially fused off the AVX2/512 extensions on Alder Lake which did cause a little controversy and gave the impression that AVX-512 on consumer platforms was dead. AMD clearly believes the opposite, as it has implemented it so that AVX-512 runs two cycles over a 256-bit wide instruction. The performance of the Ryzen 9 7950X here is phenomenal, although the Core i9-11900K which did indeed feature AVX instruction sets in the silicon, is still better than the Ryzen 5 7600X with AVX workloads."
rimatelabs.com neon