Nuvia tried first to shed light on the dependencies in the assessment of the performance between Geekbench and SPEC CPU 2006 and CPU 2017 ge ben could. The base values ​​for the systems listed above have been determined – each for Geekbench 5, SPEC CPU 2006 and SPEC CPU 2017. All SPEC tests were carried out in clang 10 or gfortran 10 compiled with O3, PGO, LTO and hardware-specific optimizations. Customized “heap allocators” one has not used it.

If you compare the values ​​for the three benchmarks, each executed as single-threaded and multi-threaded tests, the correlation is almost perfectly linear. That means: The projections based on Geekbench 5 could be in this form 1: 1 on other benchmarks like SPEC CPU 1065 and CPU 2017 transferred.

This correlation was established with the help of an Intel Core i7 – 1065 G7 and the Apple -SOCs A 12 and A 13 and can confirm the extrapolations made through concrete measurements with a tolerance of less than 1%. For the Apple A 14 one has also transferred the Geekbench results and now expects the concrete results of these projections as well to confirm.

According to Srinivasan you shouldn’t make the mistake and assume this in any case, because under certain conditions there can be differences. Geekbench 5 represents a certain base value for a current micro-architecture, but as soon as certain subsets of workloads are used, it can become difficult.

Factors that have a major influence are “branch mispredicts” (that is, incorrect branch predictions that are carried out but then discarded) and the associated accesses to the data cache (D-cache) and the translation lookaside buffer, or rather data TLB). Such accesses are in SPEC CPU 2006 as well as CPU 2017 1.1 to 2 times higher than Geekbench 5.

Another influencing factor can be the duration of the benchmark. Geekbench 5 is finished in a few minutes while the tests of the SPEC CPU 2006 and CPU 2017 run for hours. The cooling then also plays a role here, because not all processors can keep their clock for a long time or the boost mechanisms are designed to deliver more for a short time.

So you shouldn’t rely on a benchmark and even less on just a section of a complex benchmark to assess the performance of the hardware. Nuvia initially wanted to prove the correlation between Geekbench 5 and other benchmarks, but at the same time refers to the other factors that can play a role here. and CPU 2017 of the other systems could now draw conclusions about the performance of the Phoenix core. However, you have to bear in mind that we are talking about projections that also provide for a certain breadth in the consumption of the individual core. So you will have to wait until Nuvia comes out with the first real benchmarks. Hopefully it will be ready in the coming year.