You seem to imply that HPC means supercomputers and compute clusters while the fact is that many HPC workloads like modeling, simulation are also done on high-end workstations, which till only a few years ago were commonly 8-core or 16-core systems, and now 'client' systems are able to match those in core-count. So SPECfp is not irrelevant.
Sure, fair point. But if you really want "IPC" then it has to be about the robustness of the architecture when it comes to performance since we're dealing with a general purpose CPU. Performance in vectors, cryptography, and graphics, or memory bandwidth bound can be made to be boosted way above architectural changes.
That's why I disregard SpecFP. You didn't judge Pentium 4's performance entirely using it's SSE2 performance because while it was pretty fantastic, it had limited support and the performance of the x87 FPU was very relevant in those days. The overall architecture sucked.
You don't judge Rocketlake's performance using AVX512, for the same reason.
Also why do you assume 85% scaling? I know frequency scaling isn't 100% but to assign a number less than 100% also seems arbitrary.
80-85% is a common range you'll get when you isolate for variables such as compilers and Turbo modes for SpecInt suites. The FP portion highly depends on whether the system has enough memory bandwidth or not.
it's not a real up to figure, this is the point. IntelUser2000 claimed it's a real up to figure based on only one test - he made a mistake.
I believe you there, but does it really matter? I cared for Tigerlake since the iGPU there was at the top of it's game. It matters whether it's equal to competition or 20-30% above.
In this case it's 50% above the severely underpowered UHD 630 graphics. Aside from academic reasons it doesn't matter.
Just like Rocketlake is only relevant from the point that they were able to port a 10nm core to 14nm. They were able to do it. So what? It's performance and power efficiency still sucks horribly.