Does y-cruncher scale linearly with frequency, especially on high core count CPUs?
If a). it's single socket and b). the memory footprint is small enough to fit in RAM only, then generally, yes. Extremely slow RAM may hold back proper frequency scaling (especially over large frequency jumps). I'm assuming, for the sake of argument, that a hypothetical 10980XE and 3990X will have fast enough RAM that scaling to 4 GHz and 3 GHz (respectively) will not be inhibited by RAM speed.
the numberworld link also says 2x xeon 28C wins with 15s and 1x 7940x at 3,7GHz AVX512 has 25s
doesnt look like a predictable benchmark
WRT the Xeon and EPYC numbers at the top of the chart, we have no idea what clockspeeds the EPYC is running in the bench. I am skeptical of the results since there's no way a 2x 7742 system should be losing a bench like this against a pair of Xeon Golds. The 7940x is a bit of an outlier (note it's running a newer version though; it may have newer optimizations), though if you choose to use it as a baseline for examining Skylake-X/Cascade Lake-X performance, we can extrapolate 4.0 GHz AVX512 performance of a 10980XE. In that case, a 10980XE @ 4.0 GHz AVX512 would turn in a time of 17.8s - still slower than a hypothetical 3990X @ 3.0 GHz static.
edit: I wanted to add some numbers of my own to provide some clarification:
Running a 3900x @ 4.3 GHz with DDR4-3733 14-16-14-28 1T produced the following result:
Increasing core count by 50% over the listed 4.3 GHz 3700x (41.545s) produced a result that wasn't as fast as I expected, indicating that y-cruncher may be having problems scaling with core count in some scenarios (which might explain why the EPYC 7742 x2 lost to Xeon Golds and why the 7940x outran a 7980x with only a 100 MHz increase in clockspeed). Assuming proper scaling with core count, my 3900x should turn in a score of ~27s, rather than the 35s wall time (or 33s computation time) reported above. In this case, core scaling only applied at ~72% efficiency based on wall time.
If I apply the same efficiency when extrapolating the hypothetical performance of a 64c 3990X @ 3 GHz, the 3990X would complete the benchmark in ~10s (give or take) rather than the ~7s I originally predicted. That being said, scaling from 8c->64c in this benchmark may produce even less efficiency than the 72% obtained above.
edit edit: more food for thought. When running this 3900x @ "stock" (no 4.3 GHz OC) using the 250m time from the Anandtech 3950X review:
Note that AT is only using DDR4-3200 for their 3950X. But still. Core-scaling here in the 250m benchmark is basically non-existent. This result sits right in the middle of the 3950x HP and RHP results.
Facebook
Twitter