- Joined
- Mar 7, 2008
Ok, I've done some more testing, and am again reminded how much of a pain power limited running is as clocks are all over the place, making it harder to see patterns. As such I can't go into as much detail as I hope with current data.
For single thread, there may be a weak relationship with CPU clock after all. We're looking at perhaps 2% delta off ideal scaling for a 2x difference in clock change. This disappears when you run more than 1 thread.
Rocket Lake score/core/clock is about 20% higher than Comet Lake when both run at typical clocks. Before anyone says score/core/clock shouldn't matter as we normalised for clock, look at the line above. I could probably improve the peak score/core/clock some more by running lower clock, but don't intend to do that.
There is some non-ideal scaling as the thread count increases. As such, ignore previous statements about gains from HT/SMT, as it will be under-represented due to the thread scaling. I'll have to retest with disabled cores to get a better picture on this. This is in part based on my 12 core Skylake-X results, and I also used Johan45's 5950X post earlier as an additional data point, although I'm less confident in that as I have to estimate running clocks from the screenshot.
Edit: just did a run on 7920X limited to 4 cores. The effective HT scaling is ~17% which compares max & 8 thread scores relative to 1, 2 thread scores, normalised for clock/cores. I didn't use 4 cores as there was a drop in performance there. 24 thread vs 1-2 thread only implied 4% scaling. Looking at the other data, there is a noticeable drop in score when threads = cores. Similarly looking at the 5950X results, with some rough clock normalisation the 8 thread result is equivalent to about 7 cores, and the 16 thread result is about 12 cores worth. We do have the real cores in that situation so we can't say that loss is due to SMT, unless Windows scheduler is keeping it on one CCX or something like that.
Given all the above, the core/thread scaling is not as good as I thought it was initially. Cinebench still scales better across more cores/threads. Here we do have a mild diminishing return effect as thread counts increase. For modern games I think the 8 to 18 thread results are probably most interesting.
For single thread, there may be a weak relationship with CPU clock after all. We're looking at perhaps 2% delta off ideal scaling for a 2x difference in clock change. This disappears when you run more than 1 thread.
Rocket Lake score/core/clock is about 20% higher than Comet Lake when both run at typical clocks. Before anyone says score/core/clock shouldn't matter as we normalised for clock, look at the line above. I could probably improve the peak score/core/clock some more by running lower clock, but don't intend to do that.
There is some non-ideal scaling as the thread count increases. As such, ignore previous statements about gains from HT/SMT, as it will be under-represented due to the thread scaling. I'll have to retest with disabled cores to get a better picture on this. This is in part based on my 12 core Skylake-X results, and I also used Johan45's 5950X post earlier as an additional data point, although I'm less confident in that as I have to estimate running clocks from the screenshot.
Edit: just did a run on 7920X limited to 4 cores. The effective HT scaling is ~17% which compares max & 8 thread scores relative to 1, 2 thread scores, normalised for clock/cores. I didn't use 4 cores as there was a drop in performance there. 24 thread vs 1-2 thread only implied 4% scaling. Looking at the other data, there is a noticeable drop in score when threads = cores. Similarly looking at the 5950X results, with some rough clock normalisation the 8 thread result is equivalent to about 7 cores, and the 16 thread result is about 12 cores worth. We do have the real cores in that situation so we can't say that loss is due to SMT, unless Windows scheduler is keeping it on one CCX or something like that.
Given all the above, the core/thread scaling is not as good as I thought it was initially. Cinebench still scales better across more cores/threads. Here we do have a mild diminishing return effect as thread counts increase. For modern games I think the 8 to 18 thread results are probably most interesting.
