Article suggests that it's due to the relatively large difference in cache architecture.

I suspect Skymont would indeed provide double digit percentage gains given identical cache setups. However, giving Crestmont a 24 MB L3 and a 100 MHz clock speed advantage seems to be enough to cancel out Skymont’s improved architecture.

Performance-wise, Skymont seems to be at its best in high IPC workloads with a small cache footprint. For example Skymont beats Crestmont by 20.8% in 548.exchange2, a workload that fits in Zen 4’s 32 KB L1D cache.

However if a workload is really cache unfriendly, Skymont’s ability to pull more memory bandwidth can show through. I suspect that’s what happens in Y-Cruncher and 549.fotonik3d, as both are very memory bandwidth bound on other architectures. There, Skymont posts huge gains.

A long time ago, I was planning (and got some initial setup going) to use nodes in a distributed application running on a cluster of diverse architectures to benchmark performance of different types of machine to automatically look for better performance per dollar.

Ah, sounds fun and sounds like something BOINC[1] might be good for.

Perhaps one could even analyze data submitted to existing projects to analyze performance on various platforms.

[1]: https://boinc.berkeley.edu/

If Arrow Lake has Skymont attached to the ring we'll see its full performance.